[Intel-wired-lan] [PATCH iwl-net v1] i40e: fix races in PTP external timestamp work handling

[Intel-wired-lan] [PATCH iwl-net v1] i40e: fix races in PTP external timestamp work handling

[xuanqiang.luo]

From: Xuanqiang Luo <luoxuanqiang@kylinos.cn>

The time sync interrupt queues ptp_extts0_work, which reads device
registers and reports events through pf->ptp_clock.

i40e_ptp_stop() unregisters the PHC without first disabling the event
source or draining the work. The worker can therefore access the clock or
PF after either has been freed, causing a use-after-free. Disable external
timestamp events and call disable_work_sync() before unregistering the PHC.
Disabling the work also prevents an interrupt already in progress from
queuing it after the drain.

INIT_WORK() is called from i40e_ptp_set_timestamp_mode(), which runs on
reset and whenever timestamping is reconfigured. Reinitializing pending or
running work can corrupt its workqueue state. Initialize the work once from
i40e_sw_init() instead.

Also exclude reset recovery before stopping PTP so a concurrent rebuild
cannot re-enable timestamp events or recreate the PHC during removal.

Fixes: 1050713026a0 ("i40e: add support for PTP external synchronization clock")
Signed-off-by: Xuanqiang Luo <luoxuanqiang@kylinos.cn>
---
 drivers/net/ethernet/intel/i40e/i40e.h      |  1 +
 drivers/net/ethernet/intel/i40e/i40e_main.c | 17 +++++++-------
 drivers/net/ethernet/intel/i40e/i40e_ptp.c  | 25 +++++++++++++++------
 3 files changed, 28 insertions(+), 15 deletions(-)

diff --git a/drivers/net/ethernet/intel/i40e/i40e.h b/drivers/net/ethernet/intel/i40e/i40e.h
index 83e780919ac97..e2ae61bc38786 100644
--- a/drivers/net/ethernet/intel/i40e/i40e.h
+++ b/drivers/net/ethernet/intel/i40e/i40e.h
@@ -1315,6 +1315,7 @@ int i40e_ptp_hwtstamp_set(struct net_device *netdev,
 			  struct netlink_ext_ack *extack);
 void i40e_ptp_save_hw_time(struct i40e_pf *pf);
 void i40e_ptp_restore_hw_time(struct i40e_pf *pf);
+void i40e_ptp_init_work(struct i40e_pf *pf);
 void i40e_ptp_init(struct i40e_pf *pf);
 void i40e_ptp_stop(struct i40e_pf *pf);
 int i40e_ptp_alloc_pins(struct i40e_pf *pf);
diff --git a/drivers/net/ethernet/intel/i40e/i40e_main.c b/drivers/net/ethernet/intel/i40e/i40e_main.c
index a04683004a567..2909a58d57562 100644
--- a/drivers/net/ethernet/intel/i40e/i40e_main.c
+++ b/drivers/net/ethernet/intel/i40e/i40e_main.c
@@ -12819,6 +12819,7 @@ static int i40e_sw_init(struct i40e_pf *pf)
 			 "total-port-shutdown was enabled, link-down-on-close is forced on\n");
 	}
 	mutex_init(&pf->switch_mutex);
+	i40e_ptp_init_work(pf);
 
 sw_init_done:
 	return err;
@@ -16157,14 +16158,6 @@ static void i40e_remove(struct pci_dev *pdev)
 
 	i40e_devlink_unregister(pf);
 
-	i40e_dbg_pf_exit(pf);
-
-	i40e_ptp_stop(pf);
-
-	/* Disable RSS in hw */
-	i40e_write_rx_ctl(hw, I40E_PFQF_HENA(0), 0);
-	i40e_write_rx_ctl(hw, I40E_PFQF_HENA(1), 0);
-
 	/* Grab __I40E_RESET_RECOVERY_PENDING and set __I40E_IN_REMOVE
 	 * flags, once they are set, i40e_rebuild should not be called as
 	 * i40e_prep_for_reset always returns early.
@@ -16173,6 +16166,14 @@ static void i40e_remove(struct pci_dev *pdev)
 		usleep_range(1000, 2000);
 	set_bit(__I40E_IN_REMOVE, pf->state);
 
+	i40e_dbg_pf_exit(pf);
+
+	i40e_ptp_stop(pf);
+
+	/* Disable RSS in hw */
+	i40e_write_rx_ctl(hw, I40E_PFQF_HENA(0), 0);
+	i40e_write_rx_ctl(hw, I40E_PFQF_HENA(1), 0);
+
 	if (test_bit(I40E_FLAG_SRIOV_ENA, pf->flags)) {
 		set_bit(__I40E_VF_RESETS_DISABLED, pf->state);
 		i40e_free_vfs(pf);
diff --git a/drivers/net/ethernet/intel/i40e/i40e_ptp.c b/drivers/net/ethernet/intel/i40e/i40e_ptp.c
index ff62b5f2c8150..fd51ab8c10c20 100644
--- a/drivers/net/ethernet/intel/i40e/i40e_ptp.c
+++ b/drivers/net/ethernet/intel/i40e/i40e_ptp.c
@@ -169,6 +169,17 @@ static void i40e_ptp_extts0_work(struct work_struct *work)
 	ptp_clock_event(pf->ptp_clock, &event);
 }
 
+/**
+ * i40e_ptp_init_work - Initialize PTP work for a PF
+ * @pf: Board private structure
+ *
+ * Initialize work which must remain valid for the lifetime of the PF.
+ */
+void i40e_ptp_init_work(struct i40e_pf *pf)
+{
+	INIT_WORK(&pf->ptp_extts0_work, i40e_ptp_extts0_work);
+}
+
 /**
  * i40e_is_ptp_pin_dev - check if device supports PTP pins
  * @hw: pointer to the hardware structure
@@ -1186,8 +1197,6 @@ static int i40e_ptp_set_timestamp_mode(struct i40e_pf *pf,
 	regval |= 1 << I40E_PRTTSYN_CTL0_EVENT_INT_ENA_SHIFT;
 	wr32(hw, I40E_PRTTSYN_CTL0, regval);
 
-	INIT_WORK(&pf->ptp_extts0_work, i40e_ptp_extts0_work);
-
 	switch (config->tx_type) {
 	case HWTSTAMP_TX_OFF:
 		pf->ptp_tx = false;
@@ -1541,6 +1550,13 @@ void i40e_ptp_stop(struct i40e_pf *pf)
 	struct i40e_hw *hw = &pf->hw;
 	u32 regval;
 
+	/* Stop external timestamp events before unregistering the clock. */
+	regval = rd32(hw, I40E_PRTTSYN_CTL0);
+	regval &= ~I40E_PRTTSYN_CTL0_EVENT_INT_ENA_MASK;
+	wr32(hw, I40E_PRTTSYN_CTL0, regval);
+
+	disable_work_sync(&pf->ptp_extts0_work);
+
 	clear_bit(I40E_FLAG_PTP_ENA, pf->flags);
 	pf->ptp_tx = false;
 	pf->ptp_rx = false;
@@ -1570,10 +1586,5 @@ void i40e_ptp_stop(struct i40e_pf *pf)
 	regval &= ~I40E_PRTTSYN_AUX_0_PTPFLAG_MASK;
 	wr32(hw, I40E_PRTTSYN_AUX_0(0), regval);
 
-	/* Disable interrupts */
-	regval = rd32(hw, I40E_PRTTSYN_CTL0);
-	regval &= ~I40E_PRTTSYN_CTL0_EVENT_INT_ENA_MASK;
-	wr32(hw, I40E_PRTTSYN_CTL0, regval);
-
 	i40e_ptp_free_pins(pf);
 }
-- 
2.43.0

Re: [Intel-wired-lan] [PATCH iwl-net v1] i40e: fix races in PTP external timestamp work handling

[Simon Horman]

This is an AI-generated review of your patch. The human sending this
email has considered the AI review valid, or at least plausible.
Full review at: https://sashiko.dev/#/patchset/20260724093448.63154-1-xuanqiang.luo@linux.dev?part=1
---
--- Patch [1]: [PATCH iwl-net v1] i40e: fix races in PTP external timestamp work handling ---
commit 4359cf206cb2602e153fdeac3151f8d4bacf9195
Author: Xuanqiang Luo <luoxuanqiang@kylinos.cn>

i40e: fix races in PTP external timestamp work handling

This commit disables external timestamp events and drains the work
queue before unregistering the PHC to fix a use-after-free, and
moves the work initialization to device init instead of reset. It also
excludes reset recovery during teardown to avoid concurrent rebuilds.

> diff --git a/drivers/net/ethernet/intel/i40e/i40e_main.c b/drivers/net/ethernet/intel/i40e/i40e_main.c
> index a04683004a567..2909a58d57562 100644
> --- a/drivers/net/ethernet/intel/i40e/i40e_main.c
> +++ b/drivers/net/ethernet/intel/i40e/i40e_main.c
[ ... ]
> @@ -16157,14 +16158,6 @@ static void i40e_remove(struct pci_dev *pdev)
>  
>  	i40e_devlink_unregister(pf);
>  
> -	i40e_dbg_pf_exit(pf);
> -
> -	i40e_ptp_stop(pf);
> -
> -	/* Disable RSS in hw */
> -	i40e_write_rx_ctl(hw, I40E_PFQF_HENA(0), 0);
> -	i40e_write_rx_ctl(hw, I40E_PFQF_HENA(1), 0);
> -
>  	/* Grab __I40E_RESET_RECOVERY_PENDING and set __I40E_IN_REMOVE
>  	 * flags, once they are set, i40e_rebuild should not be called as
>  	 * i40e_prep_for_reset always returns early.
> @@ -16173,6 +16166,14 @@ static void i40e_remove(struct pci_dev *pdev)
>  		usleep_range(1000, 2000);

Simon says: I think the issue flagged below is somewhat orthogonal to
            this patch and can be considered in the context of possible
	    follow-up.

[Severity: High]
This is a pre-existing issue, but does this ad-hoc polling loop violate core
networking synchronization rules?

The subsystem guidelines state that using retry/poll loops on flags to guard
code sections is a bug, as it evades lockdep deadlock detection and lacks
proper memory ordering guarantees. 

i40e_remove() spins on __I40E_RESET_RECOVERY_PENDING with a sleep instead of
using a standard lock or workqueue synchronization, which exposes the teardown
path to races during hotplug or PCIe AER unbinds.

>  	set_bit(__I40E_IN_REMOVE, pf->state);
>  
> +	i40e_dbg_pf_exit(pf);
> +
> +	i40e_ptp_stop(pf);

[ ... ]
> diff --git a/drivers/net/ethernet/intel/i40e/i40e_ptp.c b/drivers/net/ethernet/intel/i40e/i40e_ptp.c
> index ff62b5f2c8150..fd51ab8c10c20 100644
> --- a/drivers/net/ethernet/intel/i40e/i40e_ptp.c
> +++ b/drivers/net/ethernet/intel/i40e/i40e_ptp.c
[ ... ]
> @@ -1541,6 +1550,13 @@ void i40e_ptp_stop(struct i40e_pf *pf)
>  	struct i40e_hw *hw = &pf->hw;
>  	u32 regval;
>  
> +	/* Stop external timestamp events before unregistering the clock. */
> +	regval = rd32(hw, I40E_PRTTSYN_CTL0);
> +	regval &= ~I40E_PRTTSYN_CTL0_EVENT_INT_ENA_MASK;
> +	wr32(hw, I40E_PRTTSYN_CTL0, regval);
> +
> +	disable_work_sync(&pf->ptp_extts0_work);
> +

[Severity: Medium]
Does this reordering introduce a TOCTOU race condition with the PTP ioctl?

The flag I40E_FLAG_PTP_ENA is still set here while we disable the workqueue.
Because i40e_remove() does not hold the rtnl_lock(), a concurrent
SIOCSHWTSTAMP ioctl could execute i40e_ptp_hwtstamp_set().

Since the ioctl checks I40E_FLAG_PTP_ENA, it might observe it as still set
and call i40e_ptp_set_timestamp_mode(). This would overwrite I40E_PRTTSYN_CTL0
and re-enable hardware interrupts immediately after i40e_ptp_stop() just
disabled them, leaving the device generating events for a disabled workqueue.

>  	clear_bit(I40E_FLAG_PTP_ENA, pf->flags);
>  	pf->ptp_tx = false;
>  	pf->ptp_rx = false;
>  
>  	if (pf->ptp_tx_skb) {
>  		struct sk_buff *skb = pf->ptp_tx_skb;
>  
>  		pf->ptp_tx_skb = NULL;
>  		clear_bit_unlock(__I40E_PTP_TX_IN_PROGRESS, pf->state);
>  		dev_kfree_skb_any(skb);
>  	}

Simon says: Although the issue flagged below is flagged as pre-existing
            I am concerned that it is exacerbated by this patch.

[Severity: High]
This is a pre-existing issue, but could this freeing pattern result in a
double-free or use-after-free?

A user configuring PTP and transmitting a packet sets pf->ptp_tx_skb.
The watchdog task in i40e_ptp_tx_hang() checks the timeout and reads
pf->ptp_tx_skb without an atomic test-and-set claim.

If a hardware interrupt arrives at this exact time, i40e_ptp_tx_hwtstamp()
can preempt the watchdog, read the exact same pointer, set it to NULL,
call skb_tstamp_tx() and then dev_kfree_skb_any().

When the interrupt returns, the watchdog task resumes and frees the exact
same SKB pointer again. This teardown path in i40e_ptp_stop() also forcibly
frees ptp_tx_skb without atomic protections against the interrupt handler,
compounding the race.

[ ... ]