Skip to contentSkip to conversation

PCI/ASPM: fix UAF by disable ASPM for link when child function is removed

PATCH Superseded Ding Hui 3 participants 5 messages May 4, 2023 – May 6, 2023 PCI

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Ding Hui

author
#0 lore raw

PATCH 1/1 PCI/ASPM: fix UAF by disable ASPM for link when child function is removed

If the Function 0 of a Multi-Function device is software removed, a freed downstream pointer will be left in struct pcie_link_state, and then when pcie_config_aspm_link() be invoked from any path, we will trigger use-after-free.

Based on the PCIe spec about ASPM Control (PCIe r6.0, sec 7.5.3.7), for Multi-Function Devices (including ARI Devices), it is recommended that software program the same value in all Functions. For ARI Devices, ASPM Control is determined solely by the setting in Function 0.

So we can just disable ASPM of the whole component if any child function is removed, the downstream pointer will be avoided from use-after-free, that will also avoid other potential corner cases.

Fixes b5a0a9b59c81 ("PCI/ASPM: Read and set up L1 substate capabilities")
---
 drivers/pci/pcie/aspm.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/pci/pcie/aspm.c +7 -8
index 66d7514ca111..1bf8306141aa 100644
@@ -1010,18 +1010,17 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
 
 	down_read(&pci_bus_sem);
 	mutex_lock(&aspm_lock);
-	/*
-	 * All PCIe functions are in one slot, remove one function will remove
-	 * the whole slot, so just wait until we are the last function left.
-	 */
-	if (!list_empty(&parent->subordinate->devices))
-		goto out;
 
 	link = parent->link_state;
 	root = link->root;
 	parent_link = link->parent;
 
-	/* All functions are removed, so just disable ASPM for the link */
+	/*
+	 * Any function is removed (including software removing), just
+	 * disable ASPM for the link, in case we can not configure the same
+	 * setting for all functions.
+	 * See PCIe r6.0, sec 7.5.3.7.
+	 */
 	pcie_config_aspm_link(link, 0);
 	list_del(&link->sibling);
 	/* Clock PM is for endpoint device */
@@ -1032,7 +1031,7 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
 		pcie_update_aspm_capable(root);
 		pcie_config_aspm_path(parent_link);
 	}
-out:
+
 	mutex_unlock(&aspm_lock);
 	up_read(&pci_bus_sem);
 }
Signature · 1 line
2.17.1
#1 lore raw

Hi,

On 5/4/23 5:34 AM, Ding Hui wrote:

Maybe you can use the following title?

"PCI/ASPM: Fix UAF by disabling ASPM for link when child function is removed

Quoted 6 lines

If the Function 0 of a Multi-Function device is software removed, a freed downstream pointer will be left in struct pcie_link_state, and then when pcie_config_aspm_link() be invoked from any path, we will trigger use-after-free.

Based on the PCIe spec about ASPM Control (PCIe r6.0, sec 7.5.3.7),

As per PCIe spec r6.0, sec 7.5.3.7, it is recommended

Quoted 10 lines

for Multi-Function Devices (including ARI Devices), it is recommended that software program the same value in all Functions. For ARI Devices, ASPM Control is determined solely by the setting in Function 0.

So we can just disable ASPM of the whole component if any child function is removed, the downstream pointer will be avoided from use-after-free, that will also avoid other potential corner cases.

Fixes b5a0a9b59c81 ("PCI/ASPM: Read and set up L1 substate capabilities")

Any bugzilla link with error log and reproduction steps?

Quoted 1 line

Suggested-by?

Quoted 19 lines
---
 drivers/pci/pcie/aspm.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/pci/pcie/aspm.c +0 -6
index 66d7514ca111..1bf8306141aa 100644
@@ -1010,18 +1010,17 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
 
 	down_read(&pci_bus_sem);
 	mutex_lock(&aspm_lock);
-	/*
-	 * All PCIe functions are in one slot, remove one function will remove
-	 * the whole slot, so just wait until we are the last function left.
-	 */
-	if (!list_empty(&parent->subordinate->devices))
-		goto out;

drivers/pci/pcie/aspm.c void pcie_aspm_exit_link_state(struct pci_dev *pdev) · line 1021

Quoted 10 lines
drivers/pci/pcie/aspm.c +4 -1
 
 	link = parent->link_state;
 	root = link->root;
 	parent_link = link->parent;
 
-	/* All functions are removed, so just disable ASPM for the link */
+	/*
+	 * Any function is removed (including software removing), just
+	 * disable ASPM for the link, in case we can not configure the same
+	 * setting for all functions.

How about following?

/*
* For any function removed, disable ASPM for the link. See PCIe r6.0,
* sec 7.7.3.7 for details.
*/

Quoted 14 lines
drivers/pci/pcie/aspm.c +3 -1
+	 * See PCIe r6.0, sec 7.5.3.7.
+	 */
 	pcie_config_aspm_link(link, 0);
 	list_del(&link->sibling);
 	/* Clock PM is for endpoint device */
@@ -1032,7 +1031,7 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
 		pcie_update_aspm_capable(root);
 		pcie_config_aspm_path(parent_link);
 	}
-out:
+
 	mutex_unlock(&aspm_lock);
 	up_read(&pci_bus_sem);
 }
Signature · 2 lines
Sathyanarayanan Kuppuswamy
Linux Kernel Developer
author
#2 lore raw
Quoted 8 lines · Sathyanarayanan Kuppuswamy

Hi,

On 5/4/23 5:34 AM, Ding Hui wrote:

Maybe you can use the following title?

"PCI/ASPM: Fix UAF by disabling ASPM for link when child function is removed

Thanks, I'll fix it in next patch.

Quoted 9 lines
Quoted 6 lines

If the Function 0 of a Multi-Function device is software removed, a freed downstream pointer will be left in struct pcie_link_state, and then when pcie_config_aspm_link() be invoked from any path, we will trigger use-after-free.

Based on the PCIe spec about ASPM Control (PCIe r6.0, sec 7.5.3.7),

As per PCIe spec r6.0, sec 7.5.3.7, it is recommended

Thanks, I'll fix it in next patch.

Quoted 4 lines
Quoted 4 lines

for Multi-Function Devices (including ARI Devices), it is recommended that software program the same value in all Functions. For ARI Devices, ASPM Control is determined solely by the setting in Function 0.

Quoted 8 lines
Quoted 6 lines

So we can just disable ASPM of the whole component if any child function is removed, the downstream pointer will be avoided from use-after-free, that will also avoid other potential corner cases.

Fixes b5a0a9b59c81 ("PCI/ASPM: Read and set up L1 substate capabilities")

Any bugzilla link with error log and reproduction steps?

Yes, I should link previous
https://lore.kernel.org/lkml/[email protected]/

Since Bjorn think the ALL details is not necessary, so I'll add the reproducer and compact result in next patch.

Quoted 4 lines
Quoted 2 lines

Suggested-by?

Sorry, I'll fix it in next patch.

Quoted 33 lines
Quoted 20 lines
---
  drivers/pci/pcie/aspm.c | 15 +++++++--------
  1 file changed, 7 insertions(+), 8 deletions(-)
drivers/pci/pcie/aspm.c +0 -6
index 66d7514ca111..1bf8306141aa 100644
@@ -1010,18 +1010,17 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
 
  	down_read(&pci_bus_sem);
  	mutex_lock(&aspm_lock);
-	/*
-	 * All PCIe functions are in one slot, remove one function will remove
-	 * the whole slot, so just wait until we are the last function left.
-	 */
-	if (!list_empty(&parent->subordinate->devices))
-		goto out;

drivers/pci/pcie/aspm.c void pcie_aspm_exit_link_state(struct pci_dev *pdev) · line 1021

Quoted 10 lines
drivers/pci/pcie/aspm.c +4 -1
 
  	link = parent->link_state;
  	root = link->root;
  	parent_link = link->parent;
 
-	/* All functions are removed, so just disable ASPM for the link */
+	/*
+	 * Any function is removed (including software removing), just
+	 * disable ASPM for the link, in case we can not configure the same
+	 * setting for all functions.

How about following?

Quoted 5 lines

/*
* For any function removed, disable ASPM for the link. See PCIe r6.0,
* sec 7.7.3.7 for details.
*/

Thanks, it's better.

Quoted 15 lines
Quoted 14 lines
drivers/pci/pcie/aspm.c +3 -1
+	 * See PCIe r6.0, sec 7.5.3.7.
+	 */
  	pcie_config_aspm_link(link, 0);
  	list_del(&link->sibling);
  	/* Clock PM is for endpoint device */
@@ -1032,7 +1031,7 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
  		pcie_update_aspm_capable(root);
  		pcie_config_aspm_path(parent_link);
  	}
-out:
+
  	mutex_unlock(&aspm_lock);
  	up_read(&pci_bus_sem);
  }
Signature · 2 lines
Thanks,
- Ding Hui
#3 lore raw
In reply to Ding Hui

drivers/pci/pcie/aspm.c void pcie_aspm_exit_link_state(struct pci_dev *pdev)

Quoted 27 lines · Ding Hui

If the Function 0 of a Multi-Function device is software removed, a freed downstream pointer will be left in struct pcie_link_state, and then when pcie_config_aspm_link() be invoked from any path, we will trigger use-after-free.

Based on the PCIe spec about ASPM Control (PCIe r6.0, sec 7.5.3.7), for Multi-Function Devices (including ARI Devices), it is recommended that software program the same value in all Functions. For ARI Devices, ASPM Control is determined solely by the setting in Function 0.

So we can just disable ASPM of the whole component if any child function is removed, the downstream pointer will be avoided from use-after-free, that will also avoid other potential corner cases.

Fixes b5a0a9b59c81 ("PCI/ASPM: Read and set up L1 substate capabilities")
---
 drivers/pci/pcie/aspm.c | 15 +++++++--------
 1 file changed, 7 insertions(+), 8 deletions(-)
drivers/pci/pcie/aspm.c +0 -0
index 66d7514ca111..1bf8306141aa 100644
@@ -1010,18 +1010,17 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)

Not directly related to your patch, but this looks racy to me:

  void pcie_aspm_exit_link_state(struct pci_dev *pdev)
  {
    struct pci_dev *parent = pdev->bus->self;
    if (!parent || !parent->link_state)
      return;

    down_read(&pci_bus_sem);
    mutex_lock(&aspm_lock);

    link = parent->link_state;
    root = link->root;
    ...
    free_link_state(link);
      link->pdev->link_state = NULL;
      kfree(link);

Since we check "parent->link_state" before acquiring the locks, I suspect that removing two functions at the same time could end up with a NULL pointer dereference:

  pcie_aspm_exit_link_state(fn 0)    pcie_aspm_exit_link_state(fn 1)
  parent = X                         parent = X
  parent->link_state != NULL         parent->link_state != NULL

acquire locks
free_link_state(link)
link->pdev->link_state = NULL # aka parent->link_state kfree(link)
release locks

                                     acquire locks
                                     link = parent->link_state # now NULL
                                     root = link->root         # NULL ptr

What do you think? I guess if this *is* a race, it should be fixed by a separate patch.

Quoted 35 lines
drivers/pci/pcie/aspm.c +7 -8
 	down_read(&pci_bus_sem);
 	mutex_lock(&aspm_lock);
-	/*
-	 * All PCIe functions are in one slot, remove one function will remove
-	 * the whole slot, so just wait until we are the last function left.
-	 */
-	if (!list_empty(&parent->subordinate->devices))
-		goto out;
 
 	link = parent->link_state;
 	root = link->root;
 	parent_link = link->parent;
 
-	/* All functions are removed, so just disable ASPM for the link */
+	/*
+	 * Any function is removed (including software removing), just
+	 * disable ASPM for the link, in case we can not configure the same
+	 * setting for all functions.
+	 * See PCIe r6.0, sec 7.5.3.7.
+	 */
 	pcie_config_aspm_link(link, 0);
 	list_del(&link->sibling);
 	/* Clock PM is for endpoint device */
@@ -1032,7 +1031,7 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
 		pcie_update_aspm_capable(root);
 		pcie_config_aspm_path(parent_link);
 	}
-out:
+
 	mutex_unlock(&aspm_lock);
 	up_read(&pci_bus_sem);
 }
Signature · 1 line
2.17.1
author
#4 lore raw
Quoted 66 lines · Bjorn Helgaas

drivers/pci/pcie/aspm.c void pcie_aspm_exit_link_state(struct pci_dev *pdev)

Quoted 27 lines · Ding Hui

If the Function 0 of a Multi-Function device is software removed, a freed downstream pointer will be left in struct pcie_link_state, and then when pcie_config_aspm_link() be invoked from any path, we will trigger use-after-free.

Based on the PCIe spec about ASPM Control (PCIe r6.0, sec 7.5.3.7), for Multi-Function Devices (including ARI Devices), it is recommended that software program the same value in all Functions. For ARI Devices, ASPM Control is determined solely by the setting in Function 0.

So we can just disable ASPM of the whole component if any child function is removed, the downstream pointer will be avoided from use-after-free, that will also avoid other potential corner cases.

Fixes b5a0a9b59c81 ("PCI/ASPM: Read and set up L1 substate capabilities")
---
  drivers/pci/pcie/aspm.c | 15 +++++++--------
  1 file changed, 7 insertions(+), 8 deletions(-)
drivers/pci/pcie/aspm.c +0 -0
index 66d7514ca111..1bf8306141aa 100644
@@ -1010,18 +1010,17 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)

Not directly related to your patch, but this looks racy to me:

   void pcie_aspm_exit_link_state(struct pci_dev *pdev)
   {
     struct pci_dev *parent = pdev->bus->self;
     if (!parent || !parent->link_state)
       return;

     down_read(&pci_bus_sem);
     mutex_lock(&aspm_lock);

     link = parent->link_state;
     root = link->root;
     ...
     free_link_state(link);
       link->pdev->link_state = NULL;
       kfree(link);

Since we check "parent->link_state" before acquiring the locks, I suspect that removing two functions at the same time could end up with a NULL pointer dereference:

   pcie_aspm_exit_link_state(fn 0)    pcie_aspm_exit_link_state(fn 1)
   parent = X                         parent = X
   parent->link_state != NULL         parent->link_state != NULL
   acquire locks
   free_link_state(link)
   link->pdev->link_state = NULL # aka parent->link_state
   kfree(link)
   release locks

                                      acquire locks
                                      link = parent->link_state # now NULL
                                      root = link->root         # NULL ptr
Quoted 3 lines

What do you think? I guess if this *is* a race, it should be fixed by a separate patch.

Maybe there's no need to worry about the race, when a pci device is removing, the pci_rescan_remove_lock should be acquired.

Anyway, double check would be safer.

Quoted 36 lines
Quoted 35 lines
drivers/pci/pcie/aspm.c +7 -8
  	down_read(&pci_bus_sem);
  	mutex_lock(&aspm_lock);
-	/*
-	 * All PCIe functions are in one slot, remove one function will remove
-	 * the whole slot, so just wait until we are the last function left.
-	 */
-	if (!list_empty(&parent->subordinate->devices))
-		goto out;
 
  	link = parent->link_state;
  	root = link->root;
  	parent_link = link->parent;
 
-	/* All functions are removed, so just disable ASPM for the link */
+	/*
+	 * Any function is removed (including software removing), just
+	 * disable ASPM for the link, in case we can not configure the same
+	 * setting for all functions.
+	 * See PCIe r6.0, sec 7.5.3.7.
+	 */
  	pcie_config_aspm_link(link, 0);
  	list_del(&link->sibling);
  	/* Clock PM is for endpoint device */
@@ -1032,7 +1031,7 @@ void pcie_aspm_exit_link_state(struct pci_dev *pdev)
  		pcie_update_aspm_capable(root);
  		pcie_config_aspm_path(parent_link);
  	}
-out:
+
  	mutex_unlock(&aspm_lock);
  	up_read(&pci_bus_sem);
  }
Signature · 1 line
2.17.1
Signature · 2 lines
Thanks,
- Ding Hui