Skip to contentSkip to conversation

i3c: master: allocate IBI workqueue with WQ_HIGHPRI

PATCH Applied Stanley Chu 4 participants 4 messages Aug 10 – Sep 17 I3C

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 i3c: master: allocate IBI workqueue with WQ_HIGHPRI

From: Stanley Chu <[email protected]>

The IBI (In-Band Interrupt) workqueue is allocated with only WQ_MEM_RECLAIM, which places IBI payload processing at normal worker priority. This is inadequate given the time-sensitive nature of IBI handling.

In the I3C protocol, when a target asserts an IBI, the SDA line is held low until the master acknowledges and completes the exchange. The IRQ handler (top half) ACKs the IBI, reads the payload, emits a STOP, and immediately queues the payload processing to the per-device ordered workqueue via i3c_master_queue_ibi() — effectively the bottom half of the IBI interrupt path.

If this workqueue worker is delayed by competing normal-priority tasks, the IBI notification reaches the client driver late. For latency-sensitive clients (e.g. sensors reporting alerts, hotplug events), this defeats the purpose of using IBI over polling. Furthermore, because the ordered workqueue serialises slots, a backlog of delayed slots can exhaust the pre-allocated IBI slot pool, causing subsequent IBIs to be dropped at the hardware level.

Add WQ_HIGHPRI to ensure IBI bottom-half work is scheduled promptly after the top-half IRQ handler enqueues it, keeping the IBI processing pipeline consistent with the interrupt-like semantics the protocol demands.

---
 drivers/i3c/master.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
drivers/i3c/master.c +2 -1
index f1be38a640ca..8fdd67a031ff 100644
@@ -3505,7 +3505,8 @@ int i3c_dev_request_ibi_locked(struct i3c_dev_desc *dev,
 	if (!ibi)
 		return -ENOMEM;
 
-	ibi->wq = alloc_ordered_workqueue(dev_name(i3cdev_to_dev(dev->dev)), WQ_MEM_RECLAIM);
+	ibi->wq = alloc_ordered_workqueue(dev_name(i3cdev_to_dev(dev->dev)),
+					  WQ_MEM_RECLAIM | WQ_HIGHPRI);
 	if (!ibi->wq) {
 		kfree(ibi);
 		return -ENOMEM;
Signature · 1 line
2.34.1
#1 lore raw
Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Race condition in i3c_dev_free_ibi_locked() allows a top-half IRQ handler to cause a use-after-free and NULL pointer dereference. --

--- Patch [1]: [PATCH v1] i3c: master: allocate IBI workqueue with WQ_HIGHPRI --- commit 9b5ec34ca4a070587e40145288be3b34ce061d58
Author: Stanley Chu <[email protected]>

i3c: master: allocate IBI workqueue with WQ_HIGHPRI

This commit adds the WQ_HIGHPRI flag to the IBI (In-Band Interrupt) workqueue allocation in the I3C master driver. This ensures that IBI bottom-half work is scheduled promptly, avoiding latency issues for sensitive clients that could occur if processed at normal worker priority.

drivers/i3c/master.c int i3c_dev_request_ibi_locked(struct i3c_dev_desc *dev, · line 3949

Quoted 11 lines
drivers/i3c/master.c +2 -1
index afcd7a21a3e66..4fa90013290f1 100644
@@ -3945,7 +3945,8 @@ int i3c_dev_request_ibi_locked(struct i3c_dev_desc *dev,
 	if (!ibi)
 		return -ENOMEM;
 
-	ibi->wq = alloc_ordered_workqueue(dev_name(i3cdev_to_dev(dev->dev)), WQ_MEM_RECLAIM);
+	ibi->wq = alloc_ordered_workqueue(dev_name(i3cdev_to_dev(dev->dev)),
+					  WQ_MEM_RECLAIM | WQ_HIGHPRI);

[Severity: High]
This isn't a bug introduced by this patch, but is there a race condition in i3c_dev_free_ibi_locked() when tearing down this workqueue?

When an I3C device driver frees its IBI (e.g. during unbind), i3c_dev_free_ibi_locked() destroys the workqueue and frees dev->ibi. However, there is no synchronize_irq() to wait for executing master controller IRQ handlers to finish.

A concurrently executing IRQ handler can read the device descriptor, be briefly delayed, and then invoke i3c_master_queue_ibi():

drivers/i3c/master.c:i3c_master_queue_ibi() {
    ...
	atomic_inc(&dev->ibi->pending_ibis);
	queue_work(dev->ibi->wq, &slot->work);
    ...
}

If this happens after i3c_dev_free_ibi_locked() has already destroyed the workqueue and freed dev->ibi:

drivers/i3c/master.c:i3c_dev_free_ibi_locked() {
    ...
	if (dev->ibi->wq) {
		destroy_workqueue(dev->ibi->wq);
		dev->ibi->wq = NULL;
	}

	kfree(dev->ibi);
	dev->ibi = NULL;
    ...
}

Could this lead to a use-after-free and NULL pointer dereference when accessing dev->ibi->wq inside queue_work()?

Quoted 3 lines
drivers/i3c/master.c +0 -0
 	if (!ibi->wq) {
 		kfree(ibi);
 		return -ENOMEM;
Signature · 6 lines
Sashiko AI review · https://sashiko.dev/#/patchset/[email protected]?part=1

-- 
linux-i3c mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-i3c
Acked-by on v1 1/1
#2 lore raw
In reply to Stanley Chu
Quoted 24 lines · Stanley Chu

From: Stanley Chu <[email protected]>

The IBI (In-Band Interrupt) workqueue is allocated with only WQ_MEM_RECLAIM, which places IBI payload processing at normal worker priority. This is inadequate given the time-sensitive nature of IBI handling.

In the I3C protocol, when a target asserts an IBI, the SDA line is held low until the master acknowledges and completes the exchange. The IRQ handler (top half) ACKs the IBI, reads the payload, emits a STOP, and immediately queues the payload processing to the per-device ordered workqueue via i3c_master_queue_ibi() — effectively the bottom half of the IBI interrupt path.

If this workqueue worker is delayed by competing normal-priority tasks, the IBI notification reaches the client driver late. For latency-sensitive clients (e.g. sensors reporting alerts, hotplug events), this defeats the purpose of using IBI over polling. Furthermore, because the ordered workqueue serialises slots, a backlog of delayed slots can exhaust the pre-allocated IBI slot pool, causing subsequent IBIs to be dropped at the hardware level.

Quoted 7 lines

Add WQ_HIGHPRI to ensure IBI bottom-half work is scheduled promptly after the top-half IRQ handler enqueues it, keeping the IBI processing pipeline consistent with the interrupt-like semantics the protocol demands.

---

maintainer Applied
#3 lore raw
In reply to Stanley Chu
Quoted 13 lines · Stanley Chu

The IBI (In-Band Interrupt) workqueue is allocated with only WQ_MEM_RECLAIM, which places IBI payload processing at normal worker priority. This is inadequate given the time-sensitive nature of IBI handling.

In the I3C protocol, when a target asserts an IBI, the SDA line is held low until the master acknowledges and completes the exchange. The IRQ handler (top half) ACKs the IBI, reads the payload, emits a STOP, and immediately queues the payload processing to the per-device ordered workqueue via i3c_master_queue_ibi() — effectively the bottom half of the IBI interrupt path.

Quoted 1 line

[...]

Applied, thanks!

[1/1] i3c: master: allocate IBI workqueue with WQ_HIGHPRI

      https://git.kernel.org/i3c/c/591980aaaa46

Best regards,

Signature · 3 lines
Alexandre Belloni, co-owner and COO, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com