Skip to contentSkip to conversation

PCI/MSI: Fix UAF in msi_capability_init

PATCH Superseded Mostafa Saleh 2 participants 5 messages Apr 29, 2024 – Jun 23, 2024 PCI

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 PCI/MSI: Fix UAF in msi_capability_init

While running Android15-6.6 and hacking on latest Qemu, I observed the following with KFENCE:

[   10.761992][   T81] pcieport 0000:00:03.0: Adding to iommu group 0
[   10.782536][   T81] pcieport 0000:00:03.0: enabling device (0000 -> 0003)
[   10.814259][   T81] ==================================================================
[   10.816534][   T81] BUG: KFENCE: use-after-free read in __pci_enable_msi_range+0x2c0/0x488
[   10.816534][   T81]
[   10.820189][   T81] Use-after-free read at 0x0000000024629571 (in kfence-#12):
[   10.822652][   T81]  __pci_enable_msi_range+0x2c0/0x488
[   10.824304][   T81]  pci_alloc_irq_vectors_affinity+0xec/0x14c
[   10.826129][   T81]  pci_alloc_irq_vectors+0x18/0x28
[   10.827606][   T81]  pcie_portdrv_probe+0x1e0/0x608
[   10.828979][   T81]  pci_device_probe+0xa8/0x174
[   10.830316][   T81]  really_probe+0x150/0x2b8
[   10.831717][   T81]  __driver_probe_device+0x7c/0x130
[   10.833368][   T81]  driver_probe_device+0x40/0x118
[   10.834899][   T81]  __device_attach_driver+0xc4/0x108
[   10.836464][   T81]  bus_for_each_drv+0x8c/0xf0
[   10.837915][   T81]  __device_attach+0xa4/0x198
[   10.839260][   T81]  device_initial_probe+0x18/0x28
[   10.840684][   T81]  bus_probe_device+0xb0/0xb4
[   10.842174][   T81]  deferred_probe_work_func+0xac/0xf8
[   10.843760][   T81]  process_one_work+0x18c/0x414
[   10.845129][   T81]  worker_thread+0x41c/0x53c
[   10.846611][   T81]  kthread+0x114/0x118
[   10.847841][   T81]  ret_from_fork+0x10/0x20
[   10.849298][   T81]
[   10.850435][   T81] kfence-#12: 0x0000000008614900-0x00000000e06c228d, size=104, cache=kmalloc-128
[   10.850435][   T81]
[   10.853455][   T81] allocated by task 81 on cpu 7 at 10.808142s:
[   10.856329][   T81]  __kmem_cache_alloc_node+0x1f0/0x2bc
[   10.857988][   T81]  kmalloc_trace+0x44/0x138
[   10.859299][   T81]  msi_alloc_desc+0x3c/0x9c
[   10.860582][   T81]  msi_domain_insert_msi_desc+0x30/0x78
[   10.862058][   T81]  msi_setup_msi_desc+0x13c/0x184
[   10.862942][   T81]  __pci_enable_msi_range+0x258/0x488
[   10.863847][   T81]  pci_alloc_irq_vectors_affinity+0xec/0x14c
[   10.864821][   T81]  pci_alloc_irq_vectors+0x18/0x28
[   10.866011][   T81]  pcie_portdrv_probe+0x1e0/0x608
[   10.867047][   T81]  pci_device_probe+0xa8/0x174
[   10.867937][   T81]  really_probe+0x150/0x2b8
[   10.868774][   T81]  __driver_probe_device+0x7c/0x130
[   10.869755][   T81]  driver_probe_device+0x40/0x118
[   10.870665][   T81]  __device_attach_driver+0xc4/0x108
[   10.871654][   T81]  bus_for_each_drv+0x8c/0xf0
[   10.872585][   T81]  __device_attach+0xa4/0x198
[   10.873581][   T81]  device_initial_probe+0x18/0x28
[   10.874645][   T81]  bus_probe_device+0xb0/0xb4
[   10.875610][   T81]  deferred_probe_work_func+0xac/0xf8
[   10.876696][   T81]  process_one_work+0x18c/0x414
[   10.877573][   T81]  worker_thread+0x41c/0x53c
[   10.878468][   T81]  kthread+0x114/0x118
[   10.879331][   T81]  ret_from_fork+0x10/0x20
[   10.880238][   T81]
[   10.880767][   T81] freed by task 81 on cpu 7 at 10.811436s:
[   10.882311][   T81]  msi_domain_free_descs+0xd4/0x10c
[   10.883318][   T81]  msi_domain_free_locked.part.0+0xc0/0x1d8
[   10.884401][   T81]  msi_domain_alloc_irqs_all_locked+0xb4/0xbc
[   10.885519][   T81]  pci_msi_setup_msi_irqs+0x30/0x4c
[   10.886549][   T81]  __pci_enable_msi_range+0x2a8/0x488
[   10.887537][   T81]  pci_alloc_irq_vectors_affinity+0xec/0x14c
[   10.888576][   T81]  pci_alloc_irq_vectors+0x18/0x28
[   10.889693][   T81]  pcie_portdrv_probe+0x1e0/0x608
[   10.890719][   T81]  pci_device_probe+0xa8/0x174
[   10.891623][   T81]  really_probe+0x150/0x2b8
[   10.892534][   T81]  __driver_probe_device+0x7c/0x130
[   10.893541][   T81]  driver_probe_device+0x40/0x118
[   10.894491][   T81]  __device_attach_driver+0xc4/0x108
[   10.895433][   T81]  bus_for_each_drv+0x8c/0xf0
[   10.896381][   T81]  __device_attach+0xa4/0x198
[   10.897359][   T81]  device_initial_probe+0x18/0x28
[   10.898401][   T81]  bus_probe_device+0xb0/0xb4
[   10.899268][   T81]  deferred_probe_work_func+0xac/0xf8
[   10.900314][   T81]  process_one_work+0x18c/0x414
[   10.901279][   T81]  worker_thread+0x41c/0x53c
[   10.902138][   T81]  kthread+0x114/0x118
[   10.902916][   T81]  ret_from_fork+0x10/0x20
[   10.903816][   T81]
[   10.904429][   T81] CPU: 7 PID: 81 Comm: kworker/u16:2 Not tainted 6.6.23mostafa+ #224
[   10.906362][   T81] Hardware name: linux,dummy-virt (DT)
[   10.907934][   T81] Workqueue: events_unbound deferred_probe_work_func
[   10.909959][   T81] ==================================================================

Looking at the upstream code, it seems to have the same issue where:

Descriptor allocation done in:
__pci_enable_msi_range

    msi_capability_init
        msi_setup_msi_desc
            msi_insert_msi_desc
                msi_domain_insert_msi_desc
                    msi_alloc_desc
                        ...

Freed in case of failure in __msi_domain_alloc_locked() __pci_enable_msi_range
msi_capability_init
pci_msi_setup_msi_irqs
msi_domain_alloc_irqs_all_locked
msi_domain_alloc_locked
__msi_domain_alloc_locked => fails msi_domain_free_locked
...

That failure would propagate back till pci_msi_setup_msi_irqs() call in msi_capability_init() which have "goto err" which would access the descriptor to clear the mask.

However, we can’t make assumptions if the descriptor is freed or not as it depends on the failure location and on MSI_FLAG_FREE_MSI_DESCS, so one simple solution is to re-read it.

Also, calling pci_free_msi_irqs() after will trigger msi_domain_free_locked() again, however it re-iterates through the xa array, which should be safe.

I hit this only once, but I can confirm the bug(and verify the fix) for the upstream kernel through stubbing return of __msi_domain_alloc_locked() to fail and with KASAN enabled.

bf6e054e0e3f ("genirq/msi: Provide msi_device_populate/destroy_sysfs()") is the first commit that introduced the free logic from the context of pci_msi_setup_msi_irqs()

---
 drivers/pci/msi/msi.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
drivers/pci/msi/msi.c +4 -1
index 682fa877478f..dd961151fecf 100644
@@ -393,7 +393,10 @@ static int msi_capability_init(struct pci_dev *dev, int nvec,
 	goto unlock;
 
 err:
-	pci_msi_unmask(entry, msi_multi_mask(entry));
+	/* Re-read the descriptor as it might have been freed */
+	entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
+	if (entry)
+		pci_msi_unmask(entry, msi_multi_mask(entry));
 	pci_free_msi_irqs(dev);
 fail:
 	dev->msi_enabled = 0;
Signature · 1 line
2.44.0.769.g3c40516874-goog
#1 lore raw

Review

Quoted 6 lines · Mostafa Saleh

err:

(quoted diff) +4 -1
-	pci_msi_unmask(entry, msi_multi_mask(entry));
+	/* Re-read the descriptor as it might have been freed */
+	entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
+	if (entry)
+		pci_msi_unmask(entry, msi_multi_mask(entry));

What unmasks the entry in the NULL case?

The mask has to be undone. So you need something like the uncompiled below.

Thanks,

tglx
---

drivers/pci/msi/msi.c +8 -2
@@ -349,7 +349,7 @@ static int msi_capability_init(struct pc
 			       struct irq_affinity *affd)
 {
 	struct irq_affinity_desc *masks = NULL;
-	struct msi_desc *entry;
+	struct msi_desc *entry, desc;
 	int ret;
 
 	/* Reject multi-MSI early on irq domain enabled architectures */
@@ -374,6 +374,12 @@ static int msi_capability_init(struct pc
 	/* All MSIs are unmasked by default; mask them all */
 	entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
 	pci_msi_mask(entry, msi_multi_mask(entry));
+	/*
+	 * Copy the MSI descriptor for the error path because
+	 * pci_msi_setup_msi_irqs() will free it for the hierarchical
+	 * interrupt domain case.
+	 */
+	memcpy(&desc, entry, sizeof(desc));
 
 	/* Configure MSI capability structure */
 	ret = pci_msi_setup_msi_irqs(dev, nvec, PCI_CAP_ID_MSI);
@@ -393,7 +399,7 @@ static int msi_capability_init(struct pc
 	goto unlock;
 
 err:
-	pci_msi_unmask(entry, msi_multi_mask(entry));
+	pci_msi_unmask(&desc, msi_multi_mask(&desc));
 	pci_free_msi_irqs(dev);
 fail:
 	dev->msi_enabled = 0;
author Tested-by
#2 lore raw

Hi Thomas,

Quoted 9 lines · Thomas Gleixner

Review

Quoted 6 lines · Mostafa Saleh

err:

(quoted diff) +4 -1
-	pci_msi_unmask(entry, msi_multi_mask(entry));
+	/* Re-read the descriptor as it might have been freed */
+	entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
+	if (entry)
+		pci_msi_unmask(entry, msi_multi_mask(entry));

What unmasks the entry in the NULL case?

Apparently nothing, I missed that. (PCI isn’t really my area, I prefer dealing with non standardised platform devices :))

Quoted 34 lines

The mask has to be undone. So you need something like the uncompiled below.

Thanks,

tglx
---

drivers/pci/msi/msi.c +7 -1
@@ -349,7 +349,7 @@ static int msi_capability_init(struct pc
 			       struct irq_affinity *affd)
 {
 	struct irq_affinity_desc *masks = NULL;
-	struct msi_desc *entry;
+	struct msi_desc *entry, desc;
 	int ret;
 
 	/* Reject multi-MSI early on irq domain enabled architectures */
@@ -374,6 +374,12 @@ static int msi_capability_init(struct pc
 	/* All MSIs are unmasked by default; mask them all */
 	entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
 	pci_msi_mask(entry, msi_multi_mask(entry));
+	/*
+	 * Copy the MSI descriptor for the error path because
+	 * pci_msi_setup_msi_irqs() will free it for the hierarchical
+	 * interrupt domain case.
+	 */
+	memcpy(&desc, entry, sizeof(desc));
 
 	/* Configure MSI capability structure */
 	ret = pci_msi_setup_msi_irqs(dev, nvec, PCI_CAP_ID_MSI);

drivers/pci/msi/msi.c static int msi_capability_init(struct pc · line 405

Quoted 8 lines
drivers/pci/msi/msi.c +1 -1
@@ -393,7 +399,7 @@ static int msi_capability_init(struct pc
 	goto unlock;
 
 err:
-	pci_msi_unmask(entry, msi_multi_mask(entry));
+	pci_msi_unmask(&desc, msi_multi_mask(&desc));
 	pci_free_msi_irqs(dev);
 fail:
 	dev->msi_enabled = 0;

I tested it with my stub, but since I didn't write the code, here is my tag, let me know if you want me to respin.

Thanks,
Mostafa

#3 lore raw

Hi Thomas,

This has been here for a while without fixing, I will respin V2 with your proposed changes,
please let me know if I should add any tag by you.

Thanks,
Mostafa

Quoted 68 lines · Mostafa Saleh

Hi Thomas,

Quoted 9 lines · Thomas Gleixner

Review

Quoted 6 lines · Mostafa Saleh

err:

(quoted diff) +4 -1
-   pci_msi_unmask(entry, msi_multi_mask(entry));
+   /* Re-read the descriptor as it might have been freed */
+   entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
+   if (entry)
+           pci_msi_unmask(entry, msi_multi_mask(entry));

What unmasks the entry in the NULL case?

Apparently nothing, I missed that. (PCI isn’t really my area, I prefer dealing with non standardised platform devices :))

Quoted 34 lines

The mask has to be undone. So you need something like the uncompiled below.

Thanks,

tglx
---

drivers/pci/msi/msi.c +7 -1
@@ -349,7 +349,7 @@ static int msi_capability_init(struct pc
                             struct irq_affinity *affd)
 {
      struct irq_affinity_desc *masks = NULL;
-     struct msi_desc *entry;
+     struct msi_desc *entry, desc;
      int ret;
 
      /* Reject multi-MSI early on irq domain enabled architectures */
@@ -374,6 +374,12 @@ static int msi_capability_init(struct pc
      /* All MSIs are unmasked by default; mask them all */
      entry = msi_first_desc(&dev->dev, MSI_DESC_ALL);
      pci_msi_mask(entry, msi_multi_mask(entry));
+     /*
+      * Copy the MSI descriptor for the error path because
+      * pci_msi_setup_msi_irqs() will free it for the hierarchical
+      * interrupt domain case.
+      */
+     memcpy(&desc, entry, sizeof(desc));
 
      /* Configure MSI capability structure */
      ret = pci_msi_setup_msi_irqs(dev, nvec, PCI_CAP_ID_MSI);

drivers/pci/msi/msi.c static int msi_capability_init(struct pc · line 405

Quoted 8 lines
drivers/pci/msi/msi.c +1 -1
@@ -393,7 +399,7 @@ static int msi_capability_init(struct pc
      goto unlock;
 
 err:
-     pci_msi_unmask(entry, msi_multi_mask(entry));
+     pci_msi_unmask(&desc, msi_multi_mask(&desc));
      pci_free_msi_irqs(dev);
 fail:
      dev->msi_enabled = 0;

I tested it with my stub, but since I didn't write the code, here is my tag, let me know if you want me to respin.

Thanks,
Mostafa

#4 lore raw
Quoted 3 lines · Mostafa Saleh

This has been here for a while without fixing, I will respin V2 with your proposed changes,
please let me know if I should add any tag by you.

Suggested-by is fine