[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20190207220247.vgzgqn627hkwnllp@gondor.apana.org.au>
Date: Fri, 8 Feb 2019 06:02:47 +0800
From: Herbert Xu <herbert@...dor.apana.org.au>
To: Johannes Berg <johannes@...solutions.net>
Cc: linux-wireless@...r.kernel.org, netdev@...r.kernel.org,
Jouni Malinen <j@...fi>, Thomas Graf <tgraf@...g.ch>
Subject: Re: [PATCH v2] rhashtable: make walk safe from softirq context
On Fri, Feb 08, 2019 at 05:48:34AM +0800, Herbert Xu wrote:
>
> > > Could you please show me who is doing this so I can review that
> > > to see whether it's a legitimate use of this API?
> >
> > I'm sure you'll say it's not legitimate, but it still exists ;-)
> >
> > mesh_plink_broken() gets called from the TX status path, via
> > ieee80211s_update_metric().
>
> Let me take a look.
OK I think it is wrong but you already knew that :)
First of all our current walk interface does not guarantee that you
will see all elements because a parallel remove can cause you to miss
elements. Although Neil Brown was trying to fix that it is still
not in the tree yet.
But more fundamentally, iterating over an unbounded list in a sotirq
context is *broken*!
Either you don't really need a hash table at all because your list
is short enough to start with, or you need to add more hash tables
(or other data structures) to efficiently look things up using these
alternative keys so that you don't end up hogging the softirq.
So this needs to be fixed in mesh.
Once that is gone hopefully we can remove the long obsolete init
API that carries the GFP flag.
Cheers,
--
Email: Herbert Xu <herbert@...dor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
Powered by blists - more mailing lists