Re: [PATCH v2 net-next] tcp: avoid expensive pskb_expand_head() calls

[Rick Jones <rick.jones2@hp.com>]

On 04/19/2012 10:25 AM, Eric Dumazet wrote:
> On Thu, 2012-04-19 at 10:20 -0700, Rick Jones wrote:
>
>> By copying them to smaller buffers? Or just by altering truesize?
>> Wasn't the whole point of fixing all the broken truesize settings to
>> accurately account for memory consumed?
>
> I checked, their truesize are OK (1024+256) for ixgbe driver.
> They could be little smaller, but not that much. (512 + 256)
>
> No, its only the sk_rcvbuf is small for a tcp sender,
> and sk_add_backlog() makes sure we dont queue more than sk_rcvbuf
> bytes in backlog.

Sounds like a variation on the theme of wildly divergent 
inbound/outbound bandwidth and constraining ACKs constraining throughput 
- only with buffer sizes.

87380 is the default SO_RCVBUF right?  That should have allowed 
87380/1280 or 68 ACKs to be queued.  Without ACK stretching from GRO 
that should have covered 68 * 2896 or 196928 bytes.  Based on your 
previous 54 usec to transmit 64 KB that would be 162+ usecs to 
accumulate those ACKs, so I guess a question becomes if TCP can be 
held-off processing ACKs for > 162 usecs, and if so and that cannot be 
changed, the autotuning will have to start increasing SO_SNDBUF 
alongside so_sndbuf even if the endpoint is not receiving.  As a 
handwave, since TCP does not know the buffer size(s) used by the driver, 
by 1 MSS for every 2 MSS it adds to SO_SNDBUF.  Or find some way to do 
it "on demand" in the about to drop path.

That or bare ACKs have to be excluded from the overhead checks somehow I 
guess, or there be more aggressive copying into smaller buffers.

Thankfully, when applications make explicit setsockopt() calls, they 
tend (ok, perhaps that is a bit of a guess) to set both SO_SNDBUF and 
SO_RCVBUF at the same time.

rick