Re: [PATCH net-next] net: netdev_alloc_skb() use build_skb()

["Michael S. Tsirkin" <mst@redhat.com>]

On Mon, Jun 04, 2012 at 09:52:59PM +0200, Eric Dumazet wrote:
> On Mon, 2012-06-04 at 22:43 +0300, Michael S. Tsirkin wrote:
> > On Mon, Jun 04, 2012 at 09:29:45PM +0200, Eric Dumazet wrote:
> > > On Mon, 2012-06-04 at 21:16 +0300, Michael S. Tsirkin wrote:
> > > 
> > > > Yes but if a tcp socket then hangs on, on one of the fragments,
> > > > while the other has been freed, the whole page is still
> > > > never reused, right?
> > > > 
> > > > Doesn't this mean truesize should be 4K?
> > > > 
> > > 
> > > Yes, or more exactly PAGE_SIZE, but then performance would really go
> > > down on machines with 64KB pages.
> > > Maybe we should make the whole frag
> > > head idea enabled only for PAGE_SIZE=4096.
> > > 
> > > Not sure we want to track precise truesize, as the minimum truesize is
> > > SKB_DATA_ALIGN(length + NET_SKB_PAD) + SKB_DATA_ALIGN(sizeof(struct
> > > skb_shared_info)) (64 + 64 + 320) = 448
> > > 
> > > Its not like buggy drivers that used truesize = length
> > > 
> > > 
> > 
> > Interesting. But where's the threshold?
> > 
> 
> It all depends on the global limit you have on your machine.
> 
> If you allow tcp memory to use 10% of ram, then a systematic x4 error
> would allow it to use 40% of ram. Mabe not enough to crash.
> 
> Now you have to find a real workload able to hit this limit for real...
> 
> But, if you "allow" a driver to claim a truesize of 1 (instead of 4096),
> you can reach the limit and OOM faster
> 
> You know, even the current page stored for each socket (sk_sndmsg_page)
> can be a problem if you setup 1.000.000 tcp sockets. That can consume
> 4GB of ram (added to inode/sockets themselves)
> This is not really taken into account right now...
> 
> 

Yes but what bugs me if the box is not under memory pressure
this overestimation limits buffers for no real gain.
How about we teach tcp to use data_len for buffer
limits normally and switch to truesize when low on memory?

-- 
MST