Skip to contentSkip to conversation

Revert "PCI: Fix reference leak in pci_register_host_bridge()"

PATCH In review Xiangwei Li 2 participants 5 messages Apr 10, 2025 – Apr 14, 2025 PCI

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Revert "PCI: Fix reference leak in pci_register_host_bridge()"

This reverts commit 804443c1f27883926de94c849d91f5b7d7d696e9.

The newly added logic incorrectly sets bus_registered to true even when device_register returns an error, this is incorrect.

When device_register fails, there is no need to release the reference count, and there are no direct error-handling operations following its execution.

Therefore, this patch is meaningless and should be reverted.

Fixes 804443c1f278 ("PCI: Fix reference leak in pci_register_host_bridge()")
---
 drivers/pci/probe.c | 9 ++-------
 1 file changed, 2 insertions(+), 7 deletions(-)
drivers/pci/probe.c +2 -7
index 364fa2a514f8..8595d41add09 100644
@@ -957,7 +957,6 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
 	resource_size_t offset, next_offset;
 	LIST_HEAD(resources);
 	struct resource *res, *next_res;
-	bool bus_registered = false;
 	char addr[64], *fmt;
 	const char *name;
 	int err;
@@ -1021,7 +1020,6 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
 	name = dev_name(&bus->dev);
 
 	err = device_register(&bus->dev);
-	bus_registered = true;
 	if (err)
 		goto unregister;
 
@@ -1110,15 +1108,12 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
 unregister:
 	put_device(&bridge->dev);
 	device_del(&bridge->dev);
+
 free:
 #ifdef CONFIG_PCI_DOMAINS_GENERIC
 	pci_bus_release_domain_nr(parent, bus->domain_nr);
 #endif
-	if (bus_registered)
-		put_device(&bus->dev);
-	else
-		kfree(bus);
-
+	kfree(bus);
 	return err;
 }
 
Signature · 1 line
2.25.1
#1 lore raw
Quoted 6 lines · Xiangwei Li

This reverts commit 804443c1f27883926de94c849d91f5b7d7d696e9.

The newly added logic incorrectly sets bus_registered to true even when device_register returns an error, this is incorrect.

When device_register fails, there is no need to release the reference count,

I think you missed some thing about device_register(). This patch is wrong.

device_register()
    -> device_initialize()
              -> kobject_init()
                    -> kobject_init_internal()
                        -> kref_init(&kobj->kref);  //set kref->refcount to 1 ^^^^^^^^^^^^^^^^^^^^^

device_register() only  fails when device_add() fails, and kerf->refcount can be 1 at this time ,  so we need to call put_deivce() to release  memory.

Quoted 7 lines

and there are no direct error-handling operations following its execution.

Therefore, this patch is meaningless and should be reverted.

Fixes 804443c1f278 ("PCI: Fix reference leak in pci_register_host_bridge()")
author
#2 lore raw
Quoted 9 lines · Su Hui
Quoted 6 lines · Xiangwei Li

This reverts commit 804443c1f27883926de94c849d91f5b7d7d696e9.

The newly added logic incorrectly sets bus_registered to true even when device_register returns an error, this is incorrect.

When device_register fails, there is no need to release the reference count,

I think you missed some thing about device_register(). This patch is wrong.

Quoted 8 lines

device_register()
    -> device_initialize()
              -> kobject_init()
                    -> kobject_init_internal() -> kref_init(&kobj->kref);  //set
kref->refcount to 1
                           ^^^^^^^^^^^^^^^^^^^^^

Sorry, I missed the initialization of refcount in device_initialize, but I’m confused about the branch logic for bus_registered. Why isn’t free(bus) executed when bus_registered == true? My understanding is that the kobject_cleanup operation triggered when refcount reaches zero does not clean up the allocated bus. Could you clarify this further?

Thanks,
Xiangwei Li

Quoted 10 lines

device_register() only  fails when device_add() fails, and kerf->refcount can be 1 at this time ,  so we need to call put_deivce() to release  memory.

Quoted 7 lines

and there are no direct error-handling operations following its execution.

Therefore, this patch is meaningless and should be reverted.

Fixes 804443c1f278 ("PCI: Fix reference leak in pci_register_host_bridge()")
#3 lore raw
Quoted 17 lines · liwei (JK)
Quoted 7 lines · Su Hui
Quoted 6 lines · Xiangwei Li

This reverts commit 804443c1f27883926de94c849d91f5b7d7d696e9.

The newly added logic incorrectly sets bus_registered to true even when device_register returns an error, this is incorrect.

When device_register fails, there is no need to release the reference count,

Quoted 9 lines

I think you missed some thing about device_register(). This patch is wrong.

device_register()
     -> device_initialize()
               -> kobject_init()
                     -> kobject_init_internal()
                         -> kref_init(&kobj->kref); //set kref->refcount to 1 ^^^^^^^^^^^^^^^^^^^^^

Quoted 6 lines

Sorry, I missed the initialization of refcount in device_initialize, but I’m confused about the branch logic for bus_registered. Why isn’t free(bus) executed when bus_registered == true? My understanding is that the kobject_cleanup operation triggered when refcount reaches zero does not clean up the allocated bus. Could you clarify this further?

1020         dev_set_name(&bus->dev, "%04x:%02x", pci_domain_nr(bus), bus->number); ^^^^^^^^^^^^^^^^^^^^
                //device name is allocated, and should be freed when device_register() is failed.

1021         name = dev_name(&bus->dev);
1022
1023         err = device_register(&bus->dev);
1024         bus_registered = true;
1025         if (err)
1026                 goto unregister;
                        [...]
1117         if (bus_registered)
1118                 put_device(&bus->dev);
                           ^^^^^^^^^^^^^^^^
                           // decrement reference count  to zero and call release_pcibus_dev() to free bus. // And call kfree_const() to free device name in kobject_cleanup().

1119         else
1120                 kfree(bus);

Commit 804443c1f278 fixes the memory leak of 'name' and consistent with the annotation of device_degister():

 '* NOTE: _Never_ directly free @dev after calling this function, even
 * if it returned an error! Always use put_device() to give up the
 * reference initialized in this function instead.'

Su Hui

author
#4 lore raw
In reply to liwei (JK)
Quoted 19 lines · liwei (JK)
Quoted 7 lines · Su Hui
Quoted 6 lines · Xiangwei Li

This reverts commit 804443c1f27883926de94c849d91f5b7d7d696e9.

The newly added logic incorrectly sets bus_registered to true even when device_register returns an error, this is incorrect.

When device_register fails, there is no need to release the reference count,

Quoted 9 lines

I think you missed some thing about device_register(). This patch is wrong.

device_register()
     -> device_initialize()
               -> kobject_init()
                     -> kobject_init_internal()
                         -> kref_init(&kobj->kref);  //set kref->refcount to 1 ^^^^^^^^^^^^^^^^^^^^^

Quoted 8 lines

Sorry, I missed the initialization of refcount in device_initialize, but I’m confused about the branch logic for bus_registered. Why isn’t free(bus) executed when bus_registered == true? My understanding is that the kobject_cleanup operation triggered when refcount reaches zero does not clean up the allocated bus. Could you clarify this further?

Thanks,
Xiangwei Li

Hi Su,

Thank you very much for your explanation. Based on your response and an analysis of the corresponding release paths, this confirms that the issue I previously raised does not exist in your patch.

Release path:
kobject_cleanup

	->device_release  // device_ktype.device_release
		->release_pcibus_dev  // pcibus_class.release_pcibus_dev
			->kfree(pci_bus)   // free(bus)

Thanks,
Xiangwei Li

Quoted 11 lines
Quoted 10 lines

device_register() only  fails when device_add() fails, and kerf->refcount can be 1 at this time ,  so we need to call put_deivce() to release  memory.

Quoted 7 lines

and there are no direct error-handling operations following its execution.

Therefore, this patch is meaningless and should be reverted.

Fixes 804443c1f278 ("PCI: Fix reference leak in pci_register_host_bridge()")