Re: [PATCH net v2 1/4] ptp: ocp: prevent timestamp rearming during detach

From
Vadim Fedorenko <vadim.fedorenko@linux.dev>
Date
2026-11-19 22:23:05
Reply to
Ahmad Byagowi, 2026-11-19
On 19/11/2026 21:19, Ahmad Byagowi wrote:
Detach disables timestamp interrupts before unregistering the PTP clock.
Until the POSIX clock is unregistered, an ioctl can rearm a channel
after that disable. The clock can then be released while the timestamp
IRQ is still registered. The core cleanup does not necessarily disable
the channel: EXTTS requests do not require a PTP_PF_EXTTS pin assignment.
I think we have to fix it other way. Basically, there are several ways
to enable EXTTS in ptp_ocp:
1. Via sma* attributes in ptp_ocp sysfs entry
2. Via pin name sysfs entry under ptp device
3. Via PTP_PIN_SETFUNC ioctl
4. Via PTP_EXTTS_REQUEST ioctl

On teardown, sma* attributes (1) are removed before disabling
interrupts. Options 2 and 3 set PTP_PF_EXTTS function and reverted
during ptp_clock_unregister(). Only option 4 does not set PTP_PF_EXTTS
function to pin, because it calls ops->enable() callback directly.
To fix it properly enable() callback has to set PTP_PF_EXTTS function
if it's requested. That will make ptp_disable_all_events() do exactly
what is has to do - clear all events generators of ptp device after
removing ioctl interface.

quoted hunk ↗ jump to hunk
Use the existing device spinlock to serialize the stopping check, PPS
subscription updates and timestamp MMIO. Set ts_stopping under that lock
before detach disables the channels, so an enable already in progress
finishes first and later enables fail with -ENODEV. Keep disables
available for PTP core cleanup.

Flush the interrupt mask writes under the lock, then drop the lock and
restore IRQ flags before pci_irq_vector(), which takes the MSI descriptor
mutex, and synchronize_irq(), which may sleep. During teardown, bypass
the PPS no-state-change shortcut: a concurrent disable
may have cleared pps_req_map and dropped the lock without yet draining
its IRQ. Detach must still disable and drain that channel before
ptp_clock_unregister().

Reported by the netdev Sashiko AI review, then checked against the PTP
and POSIX-clock paths. The source-derived userspace harness reproduces
rearming with the old source. With this fix it rejects late enables,
serializes in-flight enables, drains a mocked active IRQ and covers a
disable paused after unlock but before IRQ synchronization. It checks
that MMIO is locked and vector lookup and IRQ synchronization run unlocked
with IRQ flags restored. Shared PPS/EXTTS subscriptions and all six
channels pass with ASan/UBSan. This is not an in-kernel UAF, lockdep or
physical-card test.

Codex revised the code, tests and changelog from Vadim's review, at the
request to check the spinlock approach and rerun the removal-race tests.

Fixes: 627366c51145 ("ptp: ocp: fix resource freeing order")
Suggested-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Link: https://lore.kernel.org/r/179502617840.3136.12704245716223038161@kernel.org (local)
Link: https://lore.kernel.org/r/ad6883e5-c103-4f82-8040-e5463477ca10@linux.dev (local)
Signed-off-by: Ahmad Byagowi <redacted>
---
  drivers/ptp/ptp_ocp.c | 35 ++++++++++++++++++++++++++++-------
  1 file changed, 28 insertions(+), 7 deletions(-)
diff --git a/drivers/ptp/ptp_ocp.c b/drivers/ptp/ptp_ocp.c
index 625a85b2dc..4c98558e96 100644
--- a/drivers/ptp/ptp_ocp.c
+++ b/drivers/ptp/ptp_ocp.c
@@ -367,6 +367,8 @@ struct ptp_ocp {
  	struct pci_dev		*pdev;
  	struct device		dev;
  	spinlock_t		lock;
+	/* Protected by lock; blocks timestamp arming during teardown. */
+	bool			ts_stopping;
  	struct ocp_reg __iomem	*reg;
  	struct tod_reg __iomem	*tod;
  	struct pps_reg __iomem	*pps_to_ext;
@@ -2550,6 +2552,15 @@ ptp_ocp_ts_enable(void *priv, u32 req, bool enable)
  	struct ptp_ocp_ext_src *ext = priv;
  	struct ts_reg __iomem *reg = ext->mem;
  	struct ptp_ocp *bp = ext->bp;
+	unsigned long flags;
+	bool need_sync = false;
+	int err = 0;
+
+	spin_lock_irqsave(&bp->lock, flags);
+	if (enable && bp->ts_stopping) {
+		err = -ENODEV;
+		goto out;
+	}
  
  	if (ext == bp->pps) {
  		u32 old_map = bp->pps_req_map;
@@ -2559,9 +2570,9 @@ ptp_ocp_ts_enable(void *priv, u32 req, bool enable)
  		else
  			bp->pps_req_map &= ~req;
  
-		/* if no state change, just return */
-		if ((!!old_map ^ !!bp->pps_req_map) == 0)
-			return 0;
+		/* Teardown must drain IRQs even if another disable cleared the map. */
there was no need of this in the previous patch, why do we need it now?
quoted hunk ↗ jump to hunk
+		if ((!!old_map ^ !!bp->pps_req_map) == 0 && !bp->ts_stopping)
+			goto out;
  	}
  
  	if (enable) {
@@ -2569,16 +2580,21 @@ ptp_ocp_ts_enable(void *priv, u32 req, bool enable)
  		iowrite32(1, &reg->intr_mask);
  		iowrite32(1, &reg->intr);
  	} else {
-		int irq_vec = pci_irq_vector(bp->pdev, ext->irq_vec);
-
  		iowrite32(0, &reg->intr_mask);
  		iowrite32(0, &reg->enable);
  		ioread32(&reg->intr_mask);
+		need_sync = true;
+	}
+
+out:
+	spin_unlock_irqrestore(&bp->lock, flags);
+	if (need_sync) {
+		int irq_vec = pci_irq_vector(bp->pdev, ext->irq_vec);
+
  		if (irq_vec > 0)
  			synchronize_irq(irq_vec);
  	}
-
-	return 0;
+	return err;
  }
  
  static void
@@ -5749,12 +5765,17 @@ ptp_ocp_detach_sysfs(struct ptp_ocp *bp)
  static void
  ptp_ocp_detach(struct ptp_ocp *bp)
  {
+	unsigned long flags;
  	int i;
  
  	ptp_ocp_debugfs_remove_device(bp);
  	ptp_ocp_detach_sysfs(bp);
  	ptp_ocp_attr_group_del(bp);
  	timer_delete_sync(&bp->watchdog);
+	/* Drain in-flight enables and prevent rearming before unregister. */
+	spin_lock_irqsave(&bp->lock, flags);
+	bp->ts_stopping = true;
+	spin_unlock_irqrestore(&bp->lock, flags);
  	/* Disable interrupts on all timestampers */
  	if (bp->ts0)
  		ptp_ocp_ts_enable(bp->ts0, 0, false);
  
Thread (12 messages) 12 messages, 2 authors, last reply 52m ago
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help