[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20201018121538.usz7zzb5c7zsar42@skbuf>
Date: Sun, 18 Oct 2020 12:15:39 +0000
From: Vladimir Oltean <vladimir.oltean@....com>
To: Christian Eggers <ceggers@...i.de>
CC: "netdev@...r.kernel.org" <netdev@...r.kernel.org>,
"andrew@...n.ch" <andrew@...n.ch>,
"f.fainelli@...il.com" <f.fainelli@...il.com>,
"vivien.didelot@...il.com" <vivien.didelot@...il.com>,
"kuba@...nel.org" <kuba@...nel.org>,
Kurt Kanzenbach <kurt@...utronix.de>
Subject: Re: [RFC PATCH 02/13] net: dsa: implement a central TX reallocation
procedure
On Sun, Oct 18, 2020 at 01:59:43PM +0200, Christian Eggers wrote:
> On Sunday, 18 October 2020, 13:42:06 CEST, Vladimir Oltean wrote:
> > On Sun, Oct 18, 2020 at 12:36:03PM +0200, Christian Eggers wrote:
> > > > err = pskb_expand_head(skb, max(needed_headroom, 0),
> > > >
> > > > max(needed_tailroom, 0), GFP_ATOMIC);
> > >
> > > You may remove the second max() statement (around needed_tailroom). This
> > > would size the reallocated skb more exactly to the size actually required
> > > an may save some RAM (already tested too).
> >
> > Please explain more. needed_tailroom can be negative, why should I
> > shrink the tailroom?
> Because it will not be required anymore. This may lead to smaller memory
> allocations or the excess tailroom can be reused for headroom if needed. If
> none of both applies, the tailroom will not be changed.
Understand now, you're talking about the case where the tailroom in the
skb is already larger than what we estimate the packet will need through
its entire journey in the TX path. I still won't shrink it though, I'll
keep using the second approach you suggested.
Powered by blists - more mailing lists