[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <75a6a4bb-a8f3-4996-a157-825a5834ae52@gmail.com>
Date: Tue, 30 Oct 2018 16:06:46 -0600
From: David Ahern <dsahern@...il.com>
To: David Miller <davem@...emloft.net>, 0xeffeff@...il.com
Cc: netdev@...r.kernel.org, kuznet@....inr.ac.ru,
yoshfuji@...ux-ipv6.org
Subject: Re: [PATCH net v5] net/ipv6: Add anycast addresses to a global
hashtable
On 10/30/18 12:31 PM, David Miller wrote:
> From: Jeff Barnhill <0xeffeff@...il.com>
> Date: Tue, 30 Oct 2018 07:10:58 -0400
>
>> I originally started implementing it the way you suggested; however,
>> it seemed to complicate management of that structure because it isn't
>> currently using rcu. Also, assuming that can be worked out, where
>> would I get the net from? Would I need to store a copy in ifcaddr6,
>> or is there some way to access it during ipv6_chk_acast_addr()? It
>> seems that if I don't add a copy of net, but instead access it through
>> aca_rt(?), then freeing the ifcaddr6 memory becomes problematic
>> (detaching it from idev, while read_rcu may still be accessing it).
>> On Mon, Oct 29, 2018 at 11:32 PM David Miller <davem@...emloft.net> wrote:
>
> I don't think converting the structure over to RCU, especially because
> all of the read paths (everything leading to ipv6_chk_acast_dev()) are
> taking RCU locks already.
>
> And I cannot understand how having _two_ structures to manage a piece
> of information can be less complicated than just one.
>
> You can add a backpointer to the 'idev' in ifacaddr6 to get at the
> network namespace. You don't even need to do additional reference
> counting because the idev->ac_list is always purged before an idev
> is destroyed.
>
or make the table per namespace.
Powered by blists - more mailing lists