]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commit
workqueue: mod_delayed_work_on() shouldn't queue timer on 0 delay
authorTejun Heo <tj@kernel.org>
Sun, 2 Dec 2012 00:23:42 +0000 (16:23 -0800)
committerTejun Heo <tj@kernel.org>
Sun, 2 Dec 2012 00:43:18 +0000 (16:43 -0800)
commit8852aac25e79e38cc6529f20298eed154f60b574
treedba9304157032b33339db9c8165b3d4a5a2d05b0
parent412d32e6c98527078779e5b515823b2810e40324
workqueue: mod_delayed_work_on() shouldn't queue timer on 0 delay

8376fe22c7 ("workqueue: implement mod_delayed_work[_on]()")
implemented mod_delayed_work[_on]() using the improved
try_to_grab_pending().  The function is later used, among others, to
replace [__]candel_delayed_work() + queue_delayed_work() combinations.

Unfortunately, a delayed_work item w/ zero @delay is handled slightly
differently by mod_delayed_work_on() compared to
queue_delayed_work_on().  The latter skips timer altogether and
directly queues it using queue_work_on() while the former schedules
timer which will expire on the closest tick.  This means, when @delay
is zero, that [__]cancel_delayed_work() + queue_delayed_work_on()
makes the target item immediately executable while
mod_delayed_work_on() may induce delay of upto a full tick.

This somewhat subtle difference breaks some of the converted users.
e.g. block queue plugging uses delayed_work for deferred processing
and uses mod_delayed_work_on() when the queue needs to be immediately
unplugged.  The above problem manifested as noticeably higher number
of context switches under certain circumstances.

The difference in behavior was caused by missing special case handling
for 0 delay in mod_delayed_work_on() compared to
queue_delayed_work_on().  Joonsoo Kim posted a patch to add it -
("workqueue: optimize mod_delayed_work_on() when @delay == 0")[1].
The patch was queued for 3.8 but it was described as optimization and
I missed that it was a correctness issue.

As both queue_delayed_work_on() and mod_delayed_work_on() use
__queue_delayed_work() for queueing, it seems that the better approach
is to move the 0 delay special handling to the function instead of
duplicating it in mod_delayed_work_on().

Fix the problem by moving 0 delay special case handling from
queue_delayed_work_on() to __queue_delayed_work().  This replaces
Joonsoo's patch.

[1] http://thread.gmane.org/gmane.linux.kernel/1379011/focus=1379012

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-and-tested-by: Anders Kaseorg <andersk@MIT.EDU>
Reported-and-tested-by: Zlatko Calusic <zlatko.calusic@iskon.hr>
LKML-Reference: <alpine.DEB.2.00.1211280953350.26602@dr-wily.mit.edu>
LKML-Reference: <50A78AA9.5040904@iskon.hr>
Cc: Joonsoo Kim <js1304@gmail.com>
kernel/workqueue.c