Transmission causes wakeups by unnecessary polling.

[charles]

Downstream ticket: ​https://bugzilla.redhat.com/show_bug.cgi?id=486111

I installed F11 Beta (in a virtual machine) and updated transmission to 1.51-2
(latest Rawhide).  Powertop now reports 3-4 wake ups per second: 

14.7% (  3.5)      transmission : __mod_timer (process_timeout)
 2.1% (  0.5)      transmission : schedule_hrtimeout_range (hrtimer_wakeup)

Rawhide strace attached (strace transmission -m 2> rawhide-strace.out C-c after
~30 seconds).  

We're never going to get to zero wakeups, but there is some low-hanging fruit:

• the bound socket(s) that listen for incoming peers should use libevent to know when to accept() peers, rather than polling once per second.

• there are a couple of twice-per-second timers in peer-mgr.c that can be folded into the same timer

• the tracker timer, which fires once per second, can at a minimum be modified to fire less frequently. Some of this will be moot after #1699 but a longer interval would be a good interim step.

[charles]

patch to defer timer creation until a torrent is started

[charles]

simple one-liner to increase the tracker pulse interval from 1 sec to 1.5 sec

[charles]

Those two diffs have been committed in r8394

[charles]

better:

[root@localhost gtk]# powertop --dump 2>&1 | grep " transmission : "
   0.1% (  2.1)      transmission : __mod_timer (process_timeout) 
   0.0% (  0.5)      transmission : hrtimer_start_range_ns (hrtimer_wakeup) 

[charles]

and even better, with the net.diff patch:

[root@localhost charles]# powertop --dump 2>&1 | grep " transmission : "
   0.5% (  0.9)      transmission : __mod_timer (process_timeout) 
   0.3% (  0.5)      transmission : hrtimer_start_range_ns (hrtimer_wakeup) 
   0.0% (  0.1)      transmission : __mod_timer (death_by_timeout) 

Realistically speaking, this is about as low as we're going to get unless someone else feels strongly enough about this to dig in and hack some more. I think we're reaching the point of diminishing returns wrt code complexity vs. wakeups.

[charles]

(1) use libevent to know when to accept() incoming peers. (2) kill the port forwarding timer except for when it's needed.

[charles]

net.diff committed in r8398

[charles]

Bumping back to 1.70.

The changes in net.diff are convoluted enough that I'd rather not do them in a bugfix release.