]> err.no Git - linux-2.6/commit
[TCP]: Reduce sacked_out with reno when purging write_queue
authorIlpo Järvinen <ilpo.jarvinen@helsinki.fi>
Thu, 9 Aug 2007 11:53:36 +0000 (14:53 +0300)
committerDavid S. Miller <davem@sunset.davemloft.net>
Wed, 10 Oct 2007 23:47:58 +0000 (16:47 -0700)
commit1b6d427bb7eb69e6dc4f194a5b0f4a382a16ff82
treed67f6ea9a5f581f83b4d8228fc2964c70f940d5a
parentd02596e32925edaeccee0af8eb6c229b5615de42
[TCP]: Reduce sacked_out with reno when purging write_queue

Previously TCP had a transitional state during which reno
counted segments that are already below the current window into
sacked_out, which is now prevented. In addition, re-try now
the unconditional S+L skb catching.

This approach conservatively calls just remove_sack and leaves
reset_sack() calls alone. The best solution to the whole problem
would be to first calculate the new sacked_out fully (this patch
does not move reno_sack_reset calls from original sites and thus
does not implement this). However, that would require very
invasive change to fastretrans_alert (perhaps even slicing it to
two halves). Alternatively, all callers of tcp_packets_in_flight
(i.e., users that depend on sacked_out) should be postponed
until the new sacked_out has been calculated but it isn't any
simpler alternative.

Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@helsinki.fi>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/tcp.h
net/ipv4/tcp_input.c