Re: [RFC PATCH] e1000e: Use rtnl_lock to prevent race conditions between net and pci/pm

["David Z. Dai" <zdai@linux.vnet.ibm.com>]

On Fri, 2019-10-04 at 16:36 -0700, Alexander Duyck wrote:
> From: Alexander Duyck <alexander.h.duyck@linux.intel.com>
> 
> This patch is meant to address possible race conditions that can exist
> between network configuration and power management. A similar issue was
> fixed for igb in commit 9474933caf21 ("igb: close/suspend race in
> netif_device_detach").
> 
> In addition it consolidates the code so that the PCI error handling code
> will essentially perform the power management freeze on the device prior to
> attempting a reset, and will thaw the device afterwards if that is what it
> is planning to do. Otherwise when we call close on the interface it should
> see it is detached and not attempt to call the logic to down the interface
> and free the IRQs again.
> 
> >From what I can tell the check that was adding the check for __E1000_DOWN
> in e1000e_close was added when runtime power management was added. However
> it should not be relevant for us as we perform a call to
> pm_runtime_get_sync before we call e1000_down/free_irq so it should always
> be back up before we call into this anyway.
> 
> Signed-off-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
> ---
> 
> I'm putting this out as an RFC for now. I haven't had a chance to do much
> testing yet, but I have verified no build issues, and the driver appears
> to load, link, and pass traffic without problems.
> 
> This should address issues seen with either double freeing or never freeing
> IRQs that have been seen on this and similar drivers in the past.
> 
> I'll submit this formally after testing it over the weekend assuming there
> are no issues.
> 
>  drivers/net/ethernet/intel/e1000e/netdev.c |   33 ++++++++++++++--------------
>  1 file changed, 17 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/net/ethernet/intel/e1000e/netdev.c b/drivers/net/ethernet/intel/e1000e/netdev.c
> index d7d56e42a6aa..182a2c8f12d8 100644
> --- a/drivers/net/ethernet/intel/e1000e/netdev.c
> +++ b/drivers/net/ethernet/intel/e1000e/netdev.c
> @@ -4715,12 +4715,12 @@ int e1000e_close(struct net_device *netdev)
> 
>  	pm_runtime_get_sync(&pdev->dev);
> 
> -	if (!test_bit(__E1000_DOWN, &adapter->state)) {
> +	if (netif_device_present(netdev)) {
>  		e1000e_down(adapter, true);
>  		e1000_free_irq(adapter);
> 
>  		/* Link status message must follow this format */
> -		pr_info("%s NIC Link is Down\n", adapter->netdev->name);
> +		pr_info("%s NIC Link is Down\n", netdev->name);
>  	}
> 
>  	napi_disable(&adapter->napi);
> @@ -6299,6 +6299,7 @@ static int e1000e_pm_freeze(struct device *dev)
>  	struct net_device *netdev = dev_get_drvdata(dev);
>  	struct e1000_adapter *adapter = netdev_priv(netdev);
> 
> +	rtnl_lock();
>  	netif_device_detach(netdev);
> 
>  	if (netif_running(netdev)) {
> @@ -6313,6 +6314,8 @@ static int e1000e_pm_freeze(struct device *dev)
>  		e1000e_down(adapter, false);
>  		e1000_free_irq(adapter);
>  	}
> +	rtnl_unlock();
> +
>  	e1000e_reset_interrupt_capability(adapter);
> 
>  	/* Allow time for pending master requests to run */
> @@ -6626,27 +6629,30 @@ static int __e1000_resume(struct pci_dev *pdev)
>  	return 0;
>  }
> 
> -#ifdef CONFIG_PM_SLEEP
>  static int e1000e_pm_thaw(struct device *dev)
>  {
>  	struct net_device *netdev = dev_get_drvdata(dev);
>  	struct e1000_adapter *adapter = netdev_priv(netdev);
> +	int rc = 0;
> 
>  	e1000e_set_interrupt_capability(adapter);
> -	if (netif_running(netdev)) {
> -		u32 err = e1000_request_irq(adapter);
> 
> -		if (err)
> -			return err;
> +	rtnl_lock();
> +	if (netif_running(netdev)) {
> +		rc = e1000_request_irq(adapter);
> +		if (rc)
> +			goto err_irq;
> 
>  		e1000e_up(adapter);
>  	}
> 
>  	netif_device_attach(netdev);
> -
> -	return 0;
> +	rtnl_unlock();
> +err_irq:
> +	return rc;
>  }
> 
> +#ifdef CONFIG_PM_SLEEP
>  static int e1000e_pm_suspend(struct device *dev)
>  {
>  	struct pci_dev *pdev = to_pci_dev(dev);
> @@ -6821,13 +6827,11 @@ static pci_ers_result_t e1000_io_error_detected(struct pci_dev *pdev,
>  	struct net_device *netdev = pci_get_drvdata(pdev);
>  	struct e1000_adapter *adapter = netdev_priv(netdev);
> 
> -	netif_device_detach(netdev);
> +	e1000e_pm_freeze(&pdev->dev);
> 
>  	if (state == pci_channel_io_perm_failure)
>  		return PCI_ERS_RESULT_DISCONNECT;
> 
> -	if (netif_running(netdev))
> -		e1000e_down(adapter, true);
>  	pci_disable_device(pdev);
> 
>  	/* Request a slot slot reset. */
> @@ -6893,10 +6897,7 @@ static void e1000_io_resume(struct pci_dev *pdev)
> 
>  	e1000_init_manageability_pt(adapter);
> 
> -	if (netif_running(netdev))
> -		e1000e_up(adapter);
> -
> -	netif_device_attach(netdev);
> +	e1000e_pm_thaw(&pdev->dev);
> 
>  	/* If the controller has AMT, do not set DRV_LOAD until the interface
>  	 * is up.  For all other cases, let the f/w know that the h/w is now
In e1000e_pm_thaw(), these 2 lines need to switch order to avoid
deadlock.
from:
+       rtnl_unlock();
+err_irq:

to:
+err_irq:
+       rtnl_unlock();

I will find hardware to test this patch next week. Will update the test
result later.

Thanks! - David