PCI: pnv_php: Add null checks for OpenCAPI PHBs
Subsystems (3)
PATCH v3 1/1 PCI: pnv_php: Add null checks for OpenCAPI PHBs
For OpenCAPI phb direct slots, the .pdev for php_slots will be NULL
Various sections of the code in pnv_php can do a null dereference and crash the kernel.
Originally, the issue was hit during boot:
PowerPC PowerNV PCI Hotplug Driver version: 0.1
BUG: Kernel NULL pointer dereference at 0x00000074 Faulting instruction address: 0xc000000000b75fd0 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA PowerNV ... NIP [c000000000b75fd0] pnv_php_get_adapter_state+0x60/0x154 LR [c000000000b75fbc] pnv_php_get_adapter_state+0x4c/0x154 Call Trace: [c000c0000688f990] [c000000000b75fbc] pnv_php_get_adapter_state+0x4c/0x154 (unreliable) [c000c0000688fa20] [c000000000b78bd0] pnv_php_enable+0x94/0x378 [c000c0000688fac0] [c000000000b7912c] pnv_php_register_one.isra.0+0x11c/0x1e0
This occurs for hotplug slots on root buses where bus->self == NULL, such as OpenCAPI PHB direct slots. An added debug print (not part of this patch) confirmed it was OpenCAPI:
pnv_php: slot 'OPENCAPI-0009' has NULL pdev (bus 0009:00, parent=NO (root bus))
pnv_php: slot 'OPENCAPI-0009' dn->full_name='pciex@603a000000000', compatible='ibm,power10-pau-opencapi-pciex'
This only required null check in 'pnv_php_get_adapter_state', which caused the kernel to boot.
Even with 'pnv_php_get_adapter_state' null check, there are more possible null dereferences pointed by sashiko, including cases where userspace crashes the kernel, such as:
$ cat /sys/bus/pci/slots/*/attention ... Kernel attempted to read user page (6e) - exploit attempt? (uid: 0) BUG: Kernel NULL pointer dereference on read at 0x0000006e Faulting instruction address: 0xc000000000a83334 Oops: Kernel access of bad area, sig: 11 [#1] LE PAGE_SIZE=64K MMU=Hash SMP NR_CPUS=2048 NUMA PowerNV ... [c000000046707a20] [c000000046707b90] 0xc000000046707b90 (unreliable) [c000000046707a70] [0000000000000001] 0x1 [c000000046707ab0] [c000000000acb00c] attention_read_file+0x54/0xa8 [c000000046707b30] [c000000000abfbfc] pci_slot_attr_show+0x3c/0x58 [c000000046707b50] [c0000000008181ec] sysfs_kf_seq_show+0xd4/0x204 [c000000046707be0] [c000000000815004] kernfs_seq_show+0x44/0x58
Add null checks to prevent the null dereferences.
---
Changelog:
v3:
+ split the patch from v2 series, as it's independent
+ incorporate reviews from bjorn to improve the description
v2:
+ sashiko pointed out various pre-existing null pointer derefs, which can give access to userspace to crash the kernel, fix them
---
--- drivers/pci/hotplug/pnv_php.c | 29 +++++++++++++++++++++++------ 1 file changed, 23 insertions(+), 6 deletions(-)
drivers/pci/hotplug/pnv_php.c +23 -6
@@ -47,6 +47,9 @@ static void pnv_php_disable_irq(struct pnv_php_slot *php_slot,struct pci_dev *pdev = php_slot->pdev;u16 ctrl; +if (!pdev) +return ; +if (php_slot->irq >0 ) {pcie_capability_read_word (pdev,PCI_EXP_SLTCTL , &ctrl); ctrl &= ~(PCI_EXP_SLTCTL_HPIE |
@@ -414,7 +417,8 @@ static int pnv_php_get_adapter_state(struct hotplug_slot *slot, u8 *state)*/ ret =pnv_pci_get_presence_state (php_slot->id, &presence);if (ret >=0 ) { -if (pci_pcie_type (php_slot->pdev) ==PCI_EXP_TYPE_DOWNSTREAM && +if (php_slot->pdev && +pci_pcie_type (php_slot->pdev) ==PCI_EXP_TYPE_DOWNSTREAM && presence ==OPAL_PCI_SLOT_EMPTY ) {/* * Similar to pciehp_hpc, check whether the Link Active
@@ -442,6 +446,11 @@ static int pnv_php_get_raw_indicator_status(struct hotplug_slot *slot, u8 *statestruct pci_dev *bridge = php_slot->pdev;u16 status; +if (!bridge) { + *state =0 ; +return 0 ; + } +pcie_capability_read_word (bridge,PCI_EXP_SLTCTL , &status); *state = (status & (PCI_EXP_SLTCTL_AIC |PCI_EXP_SLTCTL_PIC )) >>6 ;return 0 ;
@@ -514,11 +523,13 @@ static int pnv_php_activate_slot(struct pnv_php_slot *php_slot,* fence / freeze. */ SLOT_WARN (php_slot,"Try %d...\n" , i +1 ); -pci_set_pcie_reset_state (php_slot->pdev, - pcie_warm_reset); -msleep (250 ); -pci_set_pcie_reset_state (php_slot->pdev, - pcie_deassert_reset); +if (php_slot->pdev) { +pci_set_pcie_reset_state (php_slot->pdev, + pcie_warm_reset); +msleep (250 ); +pci_set_pcie_reset_state (php_slot->pdev, + pcie_deassert_reset); + } ret =pnv_php_set_slot_power_state ( slot,OPAL_PCI_SLOT_POWER_ON );
@@ -911,6 +922,9 @@ pnv_php_detect_clear_suprise_removal_freeze(struct pnv_php_slot *php_slot)struct eeh_pe *pe;int i, rc; +if (!pdev) +return ; +/* * When a device is surprise removed from a downstream bridge slot, * the upstream bridge port can still end up frozen due to related EEH
@@ -1093,6 +1107,9 @@ static void pnv_php_enable_irq(struct pnv_php_slot *php_slot)struct pci_dev *pdev = php_slot->pdev;int irq, ret; +if (!pdev) +return ; +/* * The MSI/MSIx interrupt might have been occupied by other * drivers. Don't populate the surprise hotplug capability
Signature · 1 line
2.54.0
[email protected]
AI reviewThank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:
Thank you for your contribution! Sashiko AI review found 4 potential issue(s) to consider:
Pre-existing issues:
- [High] NULL pointer dereference in pnv_php_alloc_slot() during error handling.
- [High] Lockless iteration over PCI bus lists risks memory corruption.
- [High] Data race on php_slot->irq can lead to a double free_irq().
- [High] Lockless modification of php_slot->state leads to TOCTOU races. --
commit 72170284e1b2c84d6ca9a8f9f30c4d6aaa2d9c12
Author: Aditya Gupta <[email protected]>
PCI: pnv_php: Add null checks for OpenCAPI PHBs
For OpenCAPI phb direct slots, the .pdev for php_slots will be NULL. Various sections of the code in pnv_php can do a null dereference and crash the kernel. Add null checks to prevent the null dereferences.
Quoted 14 lines
drivers/pci/hotplug/pnv_php.c +3
-0@@ -47,6 +47,9 @@ static void pnv_php_disable_irq(struct pnv_php_slot *php_slot,struct pci_dev *pdev = php_slot->pdev;u16 ctrl; +if (!pdev) +return ; +if (php_slot->irq >0 ) {pcie_capability_read_word (pdev,PCI_EXP_SLTCTL , &ctrl); ctrl &= ~(PCI_EXP_SLTCTL_HPIE |
drivers/pci/hotplug/pnv_php.c
Quoted 7 lines
drivers/pci/hotplug/pnv_php.c +0
-0PCI_EXP_SLTCTL_PDCE | PCI_EXP_SLTCTL_DLLSCE);pcie_capability_write_word (pdev,PCI_EXP_SLTCTL , ctrl);free_irq (php_slot->irq, php_slot); php_slot->irq =0 ; }
[Severity: High]
This is a pre-existing issue, but can the lockless check of php_slot->irq lead to a double free?
If two threads execute pnv_php_disable_irq() concurrently (such as from multiple sysfs writes on parent and child slots), could they both observe php_slot->irq > 0 and attempt to call free_irq() for the same IRQ before it is cleared?
[ ... ]
[Severity: High]
This is a pre-existing issue, but does pnv_php_alloc_slot() crash if alloc_workqueue() fails?
drivers/pci/hotplug/pnv_php.c:pnv_php_alloc_slot() {
...
if (!php_slot->wq) {
SLOT_WARN(php_slot, "Cannot alloc workqueue\n");
...
}If SLOT_WARN is invoked here, php_slot->bus is not yet initialized. The SLOT_WARN macro evaluates to dev_warn(&(sl)->bus->dev, ...), which would take the address of dev inside a NULL bus pointer. Could this trigger a kernel Oops when dev_warn() attempts to read the device name?
[Severity: High]
This is a pre-existing issue, but is it safe to traverse bus->children and bus->slots without acquiring pci_bus_sem?
drivers/pci/hotplug/pnv_php.c:pnv_php_disable_all_irqs() {
...
/* First go down child buses */
list_for_each_entry(child_bus, &bus->children, node)
pnv_php_disable_all_irqs(child_bus);
/* Disable IRQs for all pnv_php slots on this bus */
list_for_each_entry(slot, &bus->slots, list) {
...
}Could a concurrent PCI bus rescan or device hot-unplug race with this lockless traversal, leading to list corruption or a NULL pointer dereference?
[Severity: High]
This is a pre-existing issue, but is there a race condition when checking php_slot->state locklessly?
drivers/pci/hotplug/pnv_php.c:pnv_php_enable() {
...
if (php_slot->state != PNV_PHP_STATE_REGISTERED)
return 0;
...
}And in a concurrent path:
drivers/pci/hotplug/pnv_php.c:pnv_php_unregister_one() {
...
php_slot->state = PNV_PHP_STATE_OFFLINE;
pci_hp_deregister(&php_slot->slot);
...
}If a concurrent driver unbind calls pnv_php_unregister_one(), it sets the state to PNV_PHP_STATE_OFFLINE. If an asynchronous hotplug event executing pnv_php_enable() has already passed the PNV_PHP_STATE_REGISTERED check, could it erroneously operate on a slot that is being destroyed?
Signature · 1 line
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1
Hi Aditya,
The added NULL checks for OpenCAPI/direct PHB slots make sense. The patch looks overall good to me but I have one question about below change:
drivers/pci/hotplug/pnv_php.c static int pnv_php_get_raw_indicator_status(struct hotplug_slot *slot, u8 *state · line 453
Quoted 5 lines · Aditya Gupta
drivers/pci/hotplug/pnv_php.c +5
-0+if (!bridge) { + *state =0 ; +return 0 ; + } +
Is returning success with an indicator state of 0 the expected behaviour for slots without a bridge, or would returning an error such as -EOPNOTSUPP(Operation Not Supported) or -ENODEV(No Such Device) be more appropriate?
Thanks,
~ Nikhil