When a connection that involves a pipelined transaction has been stalled (i.e.
idle with an open transaction, not an idle persistent connection) for a
second move any transactions that are pipelined after the current one onto
different connections (and implicitly close this connection when done with the
current transaction).
when it has been stalled for 10 seconds (pref configurable), cancel the current
transaction itself too - depending on its state it can hopefully be restarted
on a clean connection.
the type and state patch tries hard not to form pipelines behind resources that
could become head of line blockers. But of course that requires the ability to
predict the future, and won't be perfect.
This patch reacts to a transaction that has a large response body (defined by
either a large content-length header or actually reading a large number of
chunked bytes) by cancelling any transactions that have been pipelined down the
same connection and rescheduling them elsewhere. It also changes the type of
the connection to "solo", which prevents new transactions from being pipelined
onto this one and provides class-specific negative feedback to the pipeline
manager so that near-future requests to the same host of the same type (e.g.
general) will not be pipelined but other types (e.g. img or js/css) can still
do that.
Content-Length is ideal, because it allows us to identify the problem so early.
But even actually reading the document for a fairly long time gives it a fairly
high probability of not ending soon. (i.e. long document sizes are spread over
a larger range than small ones. duh.)
The pref network.http.pipelining.maxsize controls the threshold. I set the
default at 300KB, which is roughly the bandwidth delay product of a 2mbit 120ms
rtt connection and 1 rtt is mostly what you are giving up by canceling it on
one connection and sending it on another. (modulo maybe needing a handshake).
Backs out 952d14a9e508 (bug 707173), c170c678c9ac (bug 708305), 0a5f66d5d8e4 (bug 707662), 3204b70435fe (bug 706236) and the main landing range 48807fde0339:0bd45ead1676 (bug 528288).
Losing a TCP SYN requires a long painful (typically 3 second) delay
before being retried. This patch creates a second parallel connection
attempt for any nsHttpConnection which has not become writable before
a timeout occurs.
If you assume .5% packet loss, this converts a full 3 second delay
from a 1 in 200 event into a 1 in 40,000 event.
Whichever connection establishes itself first is used. If another one
has been started and it does connect before the one being used is
closed then the extra one is handed to the connection manager for use
by a different transaction - essentially a persistent connection with
0 previous transactions on it. (Another way to think about is
pre-fetching a 3WHS on a high latency connection).
The pref network.http.connection-retry-timeout controls the amount of
time in ms to wait for success on the initial connection before beginning
the second one. Setting it to 0 disables the parallel connection, the
default is 250.
