Re: [NET][TG3] Fwd: page allocation failure with linux 3.1.1

[Eric Dumazet <eric.dumazet@gmail.com>]

Le lundi 28 novembre 2011 à 09:16 +0200, Пламен Петров a écrit :

> Well, Eric, thanks for the explanation! I will disable TSO and will
> see how is that working out for me - if I recall correctly, I had no
> problems when TSO was off.
> 
> Sorry to bother you, if the above explanation was somewhere readily
> available, but I didn't manage to find it.

You dont bother me at all :)

As a matter of fact, your mail reminded me something I wanted to do in
the past and forgot about.

Since some devices might copy skb to a linear one in their
ndo_start_xmit(), we could set a netdev limit so that TSO can still be
used on these devices, but limiting number of frags to
2^PAGE_ALLOC_COSTLY_ORDER

(Since PAGE_ALLOC_COSTLY_ORDER == 3, thats 8 frags, instead of 16 right
now on x86)

Of course, if memory is really tight, even an ATOMIC order-3 allocation
might fail. If this happens, device limit could be dynamically
decreased. In the end, only linear skb could be built by TCP.