From: "David Laight" <redacted>
Date: Thu, 5 Jan 2012 17:27:31 -0000

If you are doing that, then the network drivers must ensure they
free such skb as soon as the transmit completes - rather than
deferring the 'end of tx' processing until some later time.

Without that the sending code can't reuse the 'page' for another
request - which is one of the things I presume this allows.

TCP breaks already if a driver defers indefinitely.  All drivers
absolutely must perform TX reclaim in a small, finite, amount of
time.

(We had issues with this on SVR4 with STREAMS buffers allocated
with esballoc()...)

This isn't the Mentat stack... so it's extremely unwise to try and
match up concerns you've had in that stack over here when evaluating
changes to Linux.