Skip to contentSkip to conversation

PCI/IOV: Fix SR-IOV locking races and AB-BA deadlock

PATCH Superseded Ionut Nechita (Wind River) 3 participants 5 messages Apr 21

Subsystems (2)

v13 of 14 Superseded Latest is v14 →

lore
Share & tools
lore
#1 lore raw

PATCH v13 1/2 PCI/IOV: Make pci_lock_rescan_remove() reentrant and protect sriov_add_vfs/sriov_del_vfs

In reply to Ionut Nechita (Wind River)

From: Ionut Nechita <[email protected]>

After reverting commit 05703271c3cd ("PCI/IOV: Add PCI rescan-remove locking when enabling/disabling SR-IOV") and moving the lock to sriov_numvfs_store(), the path through driver .remove() (e.g. rmmod, or manual unbind) that calls pci_disable_sriov() directly remains unprotected against concurrent hotplug events. This affects any SR-IOV capable driver that calls pci_disable_sriov() from its .remove() callback (i40e, ice, mlx5, bnxt, etc.).

On s390, platform-generated hot-unplug events for VFs can race with sriov_del_vfs() when a PF driver is being unloaded. The platform event handler takes pci_rescan_remove_lock, but sriov_del_vfs() does not, leading to double removal and list corruption.

We cannot use a plain mutex_lock() here because sriov_del_vfs() may also be called from paths that already hold pci_rescan_remove_lock (e.g. remove_store -> pci_stop_and_remove_bus_device_locked, or sriov_numvfs_store with the lock taken by the previous patch). Using mutex_lock() in those cases would deadlock.

Make pci_lock_rescan_remove() itself reentrant by tracking the current owner task and a recursion depth counter, as suggested by Lukas Wunner and Benjamin Block, since these recursive locking scenarios exist elsewhere in the PCI subsystem:
- If the lock is already held by the current task (owner == current):
increments the depth counter and returns without re-acquiring, avoiding deadlock.
- If the lock is held by another task: blocks until the lock is released, then records the owner and sets depth to 1.
- If the lock is not held: acquires the mutex normally.

pci_unlock_rescan_remove() decrements the depth counter and releases the mutex (clearing the owner) only when the depth reaches zero. A WARN_ON catches mismatched unlock calls from tasks that do not own the lock.

This avoids relying on mutex_get_owner(), which is not exported to modules and caused link failures for builds that inline this code outside of the core kernel image.

This approach keeps the API unchanged: callers simply pair lock/unlock calls without needing to track any return value or use separate reentrant variants.

Add pci_lock_rescan_remove()/pci_unlock_rescan_remove() calls to sriov_add_vfs() and sriov_del_vfs() to protect VF addition and removal against concurrent hotplug events.

Remove the rescan/remove locking from sriov_numvfs_store() that was introduced by commit a5338e365c45 ("PCI/IOV: Fix race between SR-IOV enable/disable and hotplug"), since the locking is now handled directly in sriov_add_vfs() and sriov_del_vfs() where it is actually needed, reducing the lock scope.

Fixes 18f9e9d150fc ("PCI/IOV: Factor out sriov_add_vfs()")
Fixes 05703271c3cd ("PCI/IOV: Add PCI rescan-remove locking when enabling/disabling SR-IOV")
Fixes a5338e365c45 ("PCI/IOV: Fix race between SR-IOV enable/disable and hotplug")
---
 drivers/pci/iov.c   |  9 +++++----
 drivers/pci/probe.c | 18 ++++++++++++++++--
 2 files changed, 21 insertions(+), 6 deletions(-)
drivers/pci/iov.c +5 -4
index 91ac4e37ecb9..7ed902539051 100644
@@ -495,9 +495,7 @@ static ssize_t sriov_numvfs_store(struct device *dev,
 
 	if (num_vfs == 0) {
 		/* disable VFs */
-		pci_lock_rescan_remove();
 		ret = pdev->driver->sriov_configure(pdev, 0);
-		pci_unlock_rescan_remove();
 		goto exit;
 	}
 
@@ -509,9 +507,7 @@ static ssize_t sriov_numvfs_store(struct device *dev,
 		goto exit;
 	}
 
-	pci_lock_rescan_remove();
 	ret = pdev->driver->sriov_configure(pdev, num_vfs);
-	pci_unlock_rescan_remove();
 	if (ret < 0)
 		goto exit;
 
@@ -633,15 +629,18 @@ static int sriov_add_vfs(struct pci_dev *dev, u16 num_vfs)
 	if (dev->no_vf_scan)
 		return 0;
 
+	pci_lock_rescan_remove();
 	for (i = 0; i < num_vfs; i++) {
 		rc = pci_iov_add_virtfn(dev, i);
 		if (rc)
 			goto failed;
 	}
+	pci_unlock_rescan_remove();
 	return 0;
 failed:
 	while (i--)
 		pci_iov_remove_virtfn(dev, i);
+	pci_unlock_rescan_remove();
 
 	return rc;
 }
@@ -766,8 +765,10 @@ static void sriov_del_vfs(struct pci_dev *dev)
 	struct pci_sriov *iov = dev->sriov;
 	int i;
 
+	pci_lock_rescan_remove();
 	for (i = 0; i < iov->num_VFs; i++)
 		pci_iov_remove_virtfn(dev, i);
+	pci_unlock_rescan_remove();
 }
 
 static void sriov_disable(struct pci_dev *dev)
drivers/pci/probe.c +16 -2
index b63cd0c310bc..0e62081bc5e5 100644
@@ -3513,16 +3513,30 @@ EXPORT_SYMBOL_GPL(pci_rescan_bus);
  * routines should always be executed under this mutex.
  */
 DEFINE_MUTEX(pci_rescan_remove_lock);
+static struct task_struct *pci_rescan_remove_owner;
+static unsigned int pci_rescan_remove_depth;
 
 void pci_lock_rescan_remove(void)
 {
-	mutex_lock(&pci_rescan_remove_lock);
+	if (pci_rescan_remove_owner == current) {
+		pci_rescan_remove_depth++;
+	} else {
+		mutex_lock(&pci_rescan_remove_lock);
+		pci_rescan_remove_owner = current;
+		pci_rescan_remove_depth = 1;
+	}
 }
 EXPORT_SYMBOL_GPL(pci_lock_rescan_remove);
 
 void pci_unlock_rescan_remove(void)
 {
-	mutex_unlock(&pci_rescan_remove_lock);
+	if (WARN_ON(pci_rescan_remove_owner != current))
+		return;
+
+	if (--pci_rescan_remove_depth == 0) {
+		pci_rescan_remove_owner = NULL;
+		mutex_unlock(&pci_rescan_remove_lock);
+	}
 }
 EXPORT_SYMBOL_GPL(pci_unlock_rescan_remove);
 
Signature · 1 line
2.53.0
#0 lore raw

v13 cover PCI/IOV: Fix SR-IOV locking races and AB-BA deadlock

From: Ionut Nechita <[email protected]>

Hi Bjorn,

This is v13 of the fix for the SR-IOV race between driver .remove() and concurrent hotplug events.

Changes since v12 (Apr 21):
- Patch 1/2: replace mutex_get_owner() with explicit owner tracking (struct task_struct * + depth counter). mutex_get_owner() is not exported outside the scheduler core and caused link failures in some build configurations:

ld: vmlinux.o: in function `pci_lock_rescan_remove':
drivers/pci/probe.c: undefined reference to `mutex_get_owner'

The new implementation records the owner on first acquisition and clears it on final release, with a WARN_ON() catching mismatched unlock calls. Semantics are unchanged.
- Patch 1/2: dropped Reviewed-by/Tested-by since the locking primitive changed; please re-review.
- Patch 2/2: unchanged from v12 (keeps R-b from Niklas Schnelle and R-b/T-b from Benjamin Block).

Changes since v11 (Mar 26):
- Patch 2/2 picked up R-b Niklas Schnelle and R-b/T-b Benjamin Block
- Rebased on linux-next (next-20260420)

Changes since v10 (Mar 18):
- Patch 2/2: added kill_device() before device_release_driver() to prevent a new driver from binding between unbind and removal, closing the TOCTOU race window identified by Benjamin Block
- Patch 1/2 unchanged from v10

Changes since v9 (Mar 10):
- NEW patch 2/2: fix AB-BA deadlock in remove_store() by calling device_release_driver() before pci_stop_and_remove_bus_device_locked(), as suggested by Benjamin Block (addresses Guenter Roeck's report)
- Patch 1/2 unchanged from v9

Changes since v8 (Mar 9):
- Added Reviewed-by from Niklas Schnelle (IBM) and Tested-by (s390)
- Added Fixes tags for the three related commits
- Removed rescan/remove locking from sriov_numvfs_store() since locking is now handled in sriov_add_vfs() and sriov_del_vfs()
- Rebased on linux-next (20260309)

The AB-BA deadlock:

CPU0 (remove_store) CPU1 (unbind_store)
-------------------- --------------------
pci_lock_rescan_remove()
device_lock()
driver .remove()
sriov_del_vfs()
pci_lock_rescan_remove() <-- WAITS
pci_stop_bus_device()
device_release_driver()
device_lock() <-- WAITS

Patch 2/2 fixes this by:
1. Marking the device as dead via kill_device() so no new driver can bind (prevents TOCTOU race between unbind and removal)
2. Calling device_release_driver() before pci_stop_and_remove_bus_device_locked(), so both paths take locks in the same order: device_lock first, then pci_rescan_remove_lock

Note: the concurrent unbind_store + hotplug-event case (where the hotplug handler takes pci_rescan_remove_lock before device_lock) remains a known limitation. This is a pre-existing issue that Benjamin Block is addressing separately in:
https://lore.kernel.org/linux-pci/354b9e4a54ced67f3c89df198041df19434fe4c8.1773235561.git.bblock@linux.ibm.com/

This race has been independently observed by multiple organizations:
- IBM (s390 platform-generated hot-unplug events racing with sriov_del_vfs during PF driver unload)
- NVIDIA (tested by Dragos Tatulea in earlier versions)
- Intel (xe driver hitting lockdep warnings and deadlocks when calling pci_disable_sriov from .remove)
- Wind River (original reporter and patch author)

Test environment:
- Tested on s390 by Benjamin Block and Niklas Schnelle (IBM)
- Tested on x86_64 with Intel and NVIDIA SR-IOV devices (earlier versions)

Based on linux-next (next-20260420).

Ionut Nechita (Wind River) (2):
PCI/IOV: Make pci_lock_rescan_remove() reentrant and protect
sriov_add_vfs/sriov_del_vfs
PCI: Fix AB-BA deadlock between device_lock and pci_rescan_remove_lock
in remove_store

 drivers/pci/iov.c       |  9 +++++----
 drivers/pci/pci-sysfs.c | 30 +++++++++++++++++++++++++++++-
 drivers/pci/probe.c     | 18 ++++++++++++++++--
 3 files changed, 50 insertions(+), 7 deletions(-)
Signature · 1 line
2.53.0
#3 lore raw

PATCH v13 2/2 PCI: Fix AB-BA deadlock between device_lock and pci_rescan_remove_lock in remove_store

From: Ionut Nechita <[email protected]>

remove_store() calls pci_stop_and_remove_bus_device_locked() which takes pci_rescan_remove_lock first, then device_lock during driver release. Meanwhile, unbind_store() takes device_lock first (via device_driver_detach), and the driver's .remove() callback may call pci_disable_sriov() -> sriov_del_vfs() -> pci_lock_rescan_remove().

This creates an AB-BA deadlock:

CPU0 (remove_store) CPU1 (unbind_store)
-------------------- --------------------
pci_lock_rescan_remove()
device_lock()
driver .remove()
sriov_del_vfs()
pci_lock_rescan_remove() <-- WAITS
pci_stop_bus_device()
device_release_driver()
device_lock() <-- WAITS

Fix this by first marking the device as dead using kill_device() to prevent any new driver from binding, then calling device_release_driver() before pci_stop_and_remove_bus_device_locked().

Marking the device dead closes the race window between unbinding and removal where a new driver could theoretically bind: once the dead flag is set, the device core will refuse any new driver probe.

After device_release_driver() returns, the driver is already unbound, so the subsequent device_release_driver() call inside pci_stop_and_remove_bus_device_locked() becomes a no-op.

---
 drivers/pci/pci-sysfs.c | 30 +++++++++++++++++++++++++++++-
 1 file changed, 29 insertions(+), 1 deletion(-)
drivers/pci/pci-sysfs.c +29 -1
index d37860841260..1426328e9f05 100644
@@ -521,8 +521,36 @@ static ssize_t remove_store(struct device *dev, struct device_attribute *attr,
 	if (kstrtoul(buf, 0, &val) < 0)
 		return -EINVAL;
 
-	if (val && device_remove_file_self(dev, attr))
+	if (val && device_remove_file_self(dev, attr)) {
+		/*
+		 * Mark the device as dead so that no new driver can bind
+		 * between the unbind and the removal below.  Once the
+		 * dead flag is set, the device core will refuse any new
+		 * driver probe.
+		 */
+		device_lock(dev);
+		kill_device(dev);
+		device_unlock(dev);
+
+		/*
+		 * Unbind the driver before removing the device to avoid
+		 * an AB-BA deadlock between device_lock and
+		 * pci_rescan_remove_lock.  Without this, remove_store
+		 * takes pci_rescan_remove_lock first (via
+		 * pci_stop_and_remove_bus_device_locked) and then
+		 * device_lock during driver release, while a concurrent
+		 * unbind_store (or sriov_numvfs_store) takes device_lock
+		 * first and then pci_rescan_remove_lock (via
+		 * sriov_del_vfs), creating a circular dependency.
+		 *
+		 * By unbinding first, the driver's .remove() callback
+		 * (including any SR-IOV VF cleanup) completes before
+		 * pci_rescan_remove_lock is acquired, ensuring both
+		 * paths take locks in the same order.
+		 */
+		device_release_driver(dev);
 		pci_stop_and_remove_bus_device_locked(to_pci_dev(dev));
+	}
 	return count;
 }
 static DEVICE_ATTR_IGNORE_LOCKDEP(remove, 0220, NULL,
Signature · 1 line
2.53.0
#4 lore raw

✗ LGCI.VerificationFailed: failure for PCI/IOV: Fix SR-IOV locking races and AB-BA deadlock

In reply to Ionut Nechita (Wind River)
== Series Details ==

== Series Details ==

Series: PCI/IOV: Fix SR-IOV locking races and AB-BA deadlock
URL : https://patchwork.freedesktop.org/series/165218/
State : failure

== Summary ==

Address '[email protected]' is not on the allowlist, which prevents CI from being triggered for this patch. If you want Intel GFX CI to accept this address, please contact the script maintainers at [email protected]. Exception occurred during validation, bailing out!

Reviewed-by Tested-by on v13 1/2
#2 lore raw

Re: [PATCH v13 1/2] PCI/IOV: Make pci_lock_rescan_remove() reentrant and protect sriov_add_vfs/sriov_del_vfs

In reply to Ionut Nechita (Wind River)

drivers/pci/probe.c EXPORT_SYMBOL_GPL(pci_rescan_bus); · line 3516

Quoted 9 lines · Ionut Nechita (Wind River)
drivers/pci/probe.c +1 -0
index b63cd0c310bc..0e62081bc5e5 100644
@@ -3513,16 +3513,30 @@ EXPORT_SYMBOL_GPL(pci_rescan_bus);
  * routines should always be executed under this mutex.
  */
 DEFINE_MUTEX(pci_rescan_remove_lock);
+static struct task_struct *pci_rescan_remove_owner;

You could make this variable const:

    static const struct task_struct *pci_rescan_remove_owner;

so it is clear that this pointer is not meant to be used to modify the task_struct in any way - just for these comparisons.

Quoted 17 lines
drivers/pci/probe.c +8 -1
+static unsigned int pci_rescan_remove_depth;
 
 void pci_lock_rescan_remove(void)
 {
-	mutex_lock(&pci_rescan_remove_lock);
+	if (pci_rescan_remove_owner == current) {
+		pci_rescan_remove_depth++;
+	} else {
+		mutex_lock(&pci_rescan_remove_lock);
+		pci_rescan_remove_owner = current;
+		pci_rescan_remove_depth = 1;
+	}
 }
 EXPORT_SYMBOL_GPL(pci_lock_rescan_remove);
 
 void pci_unlock_rescan_remove(void)
 {

drivers/pci/probe.c EXPORT_SYMBOL_GPL(pci_rescan_bus); · line 3541

Quoted 10 lines
drivers/pci/probe.c +7 -1
-	mutex_unlock(&pci_rescan_remove_lock);
+	if (WARN_ON(pci_rescan_remove_owner != current))
+		return;
+
+	if (--pci_rescan_remove_depth == 0) {
+		pci_rescan_remove_owner = NULL;
+		mutex_unlock(&pci_rescan_remove_lock);
+	}
 }
 EXPORT_SYMBOL_GPL(pci_unlock_rescan_remove);

But anyway, with, or without this change, this looks good to me. I hadn't run into any linker issues, but I only ever used our debug config. Also holds up with my tests on my s390 rig.

Signature · 4 lines
Best Regards, Benjamin Block        /        Linux on IBM Z Kernel Development
IBM Deutschland Research & Development GmbH    /   https://www.ibm.com/privacy
Vors. Aufs.-R.: Wolfgang Wendt         /        Geschäftsführung: David Faller
Sitz der Ges.: Ehningen     /     Registergericht: AmtsG Stuttgart, HRB 243294