Re: [PATCH net 1/3] net: hibmcge: fix rtnl deadlock issue

[Jijie Shao <shaojijie@huawei.com>]

on 2025/8/1 18:05, Simon Horman wrote:
> On Thu, Jul 31, 2025 at 09:47:47PM +0800, Jijie Shao wrote:
>> Currently, the hibmcge netdev acquires the rtnl_lock in
>> pci_error_handlers.reset_prepare() and releases it in
>> pci_error_handlers.reset_done().
>>
>> However, in the PCI framework:
>> pci_reset_bus - __pci_reset_slot - pci_slot_save_and_disable_locked -
>>   pci_dev_save_and_disable - err_handler->reset_prepare(dev);
>>
>> In pci_slot_save_and_disable_locked():
>> 	list_for_each_entry(dev, &slot->bus->devices, bus_list) {
>> 		if (!dev->slot || dev->slot!= slot)
>> 			continue;
>> 		pci_dev_save_and_disable(dev);
>> 		if (dev->subordinate)
>> 			pci_bus_save_and_disable_locked(dev->subordinate);
>> 	}
>>
>> This will iterate through all devices under the current bus and execute
>> err_handler->reset_prepare(), causing two devices of the hibmcge driver
>> to sequentially request the rtnl_lock, leading to a deadlock.
>>
>> Since the driver now executes netif_device_detach()
>> before the reset process, it will not concurrently with
>> other netdev APIs, so there is no need to hold the rtnl_lock now.
>>
>> Therefore, this patch removes the rtnl_lock during the reset process and
>> adjusts the position of HBG_NIC_STATE_RESETTING to ensure
>> that multiple resets are not executed concurrently.
>>
>> Fixes: 3f5a61f6d504f ("net: hibmcge: Add reset supported in this module")
>> Signed-off-by: Jijie Shao <shaojijie@huawei.com>
>> ---
>>   drivers/net/ethernet/hisilicon/hibmcge/hbg_err.c | 13 ++++---------
>>   1 file changed, 4 insertions(+), 9 deletions(-)
>>
>> diff --git a/drivers/net/ethernet/hisilicon/hibmcge/hbg_err.c b/drivers/net/ethernet/hisilicon/hibmcge/hbg_err.c
>> index 503cfbfb4a8a..94bc6f0da912 100644
>> --- a/drivers/net/ethernet/hisilicon/hibmcge/hbg_err.c
>> +++ b/drivers/net/ethernet/hisilicon/hibmcge/hbg_err.c
>> @@ -53,9 +53,11 @@ static int hbg_reset_prepare(struct hbg_priv *priv, enum hbg_reset_type type)
>>   {
>>   	int ret;
>>   
>> -	ASSERT_RTNL();
>> +	if (test_and_set_bit(HBG_NIC_STATE_RESETTING, &priv->state))
>> +		return -EBUSY;
>>   
>>   	if (netif_running(priv->netdev)) {
>> +		clear_bit(HBG_NIC_STATE_RESETTING, &priv->state);
>>   		dev_warn(&priv->pdev->dev,
>>   			 "failed to reset because port is up\n");
>>   		return -EBUSY;
>> @@ -64,7 +66,6 @@ static int hbg_reset_prepare(struct hbg_priv *priv, enum hbg_reset_type type)
>>   	netif_device_detach(priv->netdev);
>>   
>>   	priv->reset_type = type;
>> -	set_bit(HBG_NIC_STATE_RESETTING, &priv->state);
>>   	clear_bit(HBG_NIC_STATE_RESET_FAIL, &priv->state);
>>   	ret = hbg_hw_event_notify(priv, HBG_HW_EVENT_RESET);
>>   	if (ret) {
>> @@ -84,10 +85,8 @@ static int hbg_reset_done(struct hbg_priv *priv, enum hbg_reset_type type)
>>   	    type != priv->reset_type)
>>   		return 0;
>>   
>> -	ASSERT_RTNL();
>> -
>> -	clear_bit(HBG_NIC_STATE_RESETTING, &priv->state);
>>   	ret = hbg_rebuild(priv);
>> +	clear_bit(HBG_NIC_STATE_RESETTING, &priv->state);
> Hi Jijie,
>
> If I understand things correctly, then with this patch the
> HBG_NIC_STATE_RESETTING bit is used to prevent concurrent execution.
>
> Noting that a reset may be triggered via eththool, where hbg_reset() is
> used as a callback, I am concerned about concurrency implications for lines
> below this one.

Yes, just like the following, it can lead to reset and net open concurrency.
===========

      reset1                                              reset2                               open

set_bit HBG_NIC_STATE_RESETTING

      netif_device_detach()
      
      resetting...

clear_bit HBG_NIC_STATE_RESETTING
                                               set_bit HBG_NIC_STATE_RESETTING
                                                    netif_device_detach()

       netif_device_attach()
                                                          resetting...                     hbg_net_open()
                                                                                           hbg_txrx_init()

                                               clear_bit HBG_NIC_STATE_RESETTING
                                                       netif_device_attach()

============
Thank you for your reminder.
I will fix it in V2

Jijie Shao

>>   	if (ret) {
>>   		priv->stats.reset_fail_cnt++;
>>   		set_bit(HBG_NIC_STATE_RESET_FAIL, &priv->state);
>> @@ -101,12 +100,10 @@ static int hbg_reset_done(struct hbg_priv *priv, enum hbg_reset_type type)
>>   	return ret;
>>   }
>>   
>> -/* must be protected by rtnl lock */
>>   int hbg_reset(struct hbg_priv *priv)
>>   {
>>   	int ret;
>>   
>> -	ASSERT_RTNL();
>>   	ret = hbg_reset_prepare(priv, HBG_RESET_TYPE_FUNCTION);
>>   	if (ret)
>>   		return ret;
>> @@ -171,7 +168,6 @@ static void hbg_pci_err_reset_prepare(struct pci_dev *pdev)
>>   	struct net_device *netdev = pci_get_drvdata(pdev);
>>   	struct hbg_priv *priv = netdev_priv(netdev);
>>   
>> -	rtnl_lock();
>>   	hbg_reset_prepare(priv, HBG_RESET_TYPE_FLR);
>>   }
>>   
>> @@ -181,7 +177,6 @@ static void hbg_pci_err_reset_done(struct pci_dev *pdev)
>>   	struct hbg_priv *priv = netdev_priv(netdev);
>>   
>>   	hbg_reset_done(priv, HBG_RESET_TYPE_FLR);
>> -	rtnl_unlock();
>>   }
>>   
>>   static const struct pci_error_handlers hbg_pci_err_handler = {
>> -- 
>> 2.33.0
>>