Re: [PATCH v2 net-next] net: link_watch: prevent starvation when processing linkwatch wq

[Yunsheng Lin <linyunsheng@huawei.com>]

On 2019/5/31 19:17, Salil Mehta wrote:
>> From: netdev-owner@vger.kernel.org [mailto:netdev-
>> owner@vger.kernel.org] On Behalf Of Yunsheng Lin
>> Sent: Friday, May 31, 2019 10:01 AM
>> To: davem@davemloft.net
>> Cc: hkallweit1@gmail.com; f.fainelli@gmail.com;
>> stephen@networkplumber.org; netdev@vger.kernel.org; linux-
>> kernel@vger.kernel.org; Linuxarm <linuxarm@huawei.com>
>> Subject: [PATCH v2 net-next] net: link_watch: prevent starvation when
>> processing linkwatch wq
>>
>> When user has configured a large number of virtual netdev, such
>> as 4K vlans, the carrier on/off operation of the real netdev
>> will also cause it's virtual netdev's link state to be processed
>> in linkwatch. Currently, the processing is done in a work queue,
>> which may cause cpu and rtnl locking starvation problem.
>>
>> This patch releases the cpu and rtnl lock when link watch worker
>> has processed a fixed number of netdev' link watch event.
>>
>> Currently __linkwatch_run_queue is called with rtnl lock, so
>> enfore it with ASSERT_RTNL();
>>
>> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
>> ---
>> V2: use cond_resched and rtnl_unlock after processing a fixed
>>     number of events
>> ---
>>  net/core/link_watch.c | 17 +++++++++++++++++
>>  1 file changed, 17 insertions(+)
>>
>> diff --git a/net/core/link_watch.c b/net/core/link_watch.c
>> index 7f51efb..07eebfb 100644
>> --- a/net/core/link_watch.c
>> +++ b/net/core/link_watch.c
>> @@ -168,9 +168,18 @@ static void linkwatch_do_dev(struct net_device
>> *dev)
>>
>>  static void __linkwatch_run_queue(int urgent_only)
>>  {
>> +#define MAX_DO_DEV_PER_LOOP	100
>> +
>> +	int do_dev = MAX_DO_DEV_PER_LOOP;
>>  	struct net_device *dev;
>>  	LIST_HEAD(wrk);
>>
>> +	ASSERT_RTNL();
>> +
>> +	/* Give urgent case more budget */
>> +	if (urgent_only)
>> +		do_dev += MAX_DO_DEV_PER_LOOP;
>> +
>>  	/*
>>  	 * Limit the number of linkwatch events to one
>>  	 * per second so that a runaway driver does not
>> @@ -200,6 +209,14 @@ static void __linkwatch_run_queue(int urgent_only)
>>  		}
>>  		spin_unlock_irq(&lweventlist_lock);
>>  		linkwatch_do_dev(dev);
>> +
>> +		if (--do_dev < 0) {
>> +			rtnl_unlock();
>> +			cond_resched();
> 
> 
> 
> Sorry, missed in my earlier comment. I could see multiple problems here
> and please correct me if I am wrong:
> 
> 1. It looks like releasing the rtnl_lock here and then res-scheduling might
>    not be safe, especially when you have already held *lweventlist_lock*
>    (which is global and not per-netdev), and when you are trying to
>    reschedule. This can cause *deadlock* with itself.
> 
>    Reason: once you release the rtnl_lock() the similar leg of function 
>    netdev_wait_allrefs() could be called for some other netdevice which
>    might end up in waiting for same global linkwatch event list lock
>    i.e. *lweventlist_lock*.

lweventlist_lock has been released before releasing the rtnl_lock and
rescheduling.

> 
> 2. After releasing the rtnl_lock() we have not ensured that all the rcu
>    operations are complete. Perhaps we need to take rcu_barrier() before
>    retaking the rtnl_lock()
Why do we need to ensure all the rcu operations are complete here?

> 
> 
> 
> 
>> +			do_dev = MAX_DO_DEV_PER_LOOP;
> 
> 
> 
> Here, I think rcu_barrier() should exist.

In netdev_wait_allrefs, rcu_barrier is indeed called between
__rtnl_unlock and rtnl_lock and is added by below commit
0115e8e30d6f ("net: remove delay at device dismantle"), which
seems to work with NETDEV_UNREGISTER_FINAL.

And the NETDEV_UNREGISTER_FINAL is removed by commit
070f2d7e264a ("net: Drop NETDEV_UNREGISTER_FINAL"), which says
something about whether the rcu_barrier is still needed.

"dev_change_net_namespace() and netdev_wait_allrefs()
have rcu_barrier() before NETDEV_UNREGISTER_FINAL call,
and the source commits say they were introduced to
delemit the call with NETDEV_UNREGISTER, but this patch
leaves them on the places, since they require additional
analysis, whether we need in them for something else."

So the reason of calling rcu_barrier in netdev_wait_allrefs
is unclear now.

Also rcu_barrier in netdev_wait_allrefs is added to fix the
device dismantle problem, so for linkwatch, maybe it is not
needed.

> 
> 
> 
>> +			rtnl_lock();
>> +		}
>> +
>>  		spin_lock_irq(&lweventlist_lock);
>>  	}
> 
> 
> .
>