| Age | Commit message (Collapse) | Author |
|---|
|
|
|
|
|
Introducing the throttling unfortunately has a negative side-effect: it
delays all stdin processing, including EOF detection, which can cause
confusion the previous commit tries to fix.
The only benefit of the throttling is to prevent 100% CPU usage when a
lot of garbage is provided on xmobar stdin. We still don't know where
that garbage comes from in https://github.com/jaor/xmobar/issues/438, or
rather why there's more than a handful of lines of such garbage. @psibi
has since fixed his setup to not produce that garbage, and no one else
ever reported such a problem, so it's probably safe to ignore it for
now.
Should anyone ever encounter that again, feel free to ping me, even in
the middle of the night, to help debug this.
Fixes: 7759df11f746 ("StdinReader: Improve exception handling")
Fixes: b7a3d6745817 ("Avoid busy looping by not catching all exceptions")
|
|
This corrects a misleading comment "EOF check is necessary for certain
systems" which was added without complete understanding of the root
cause of #442. That issue was in fact caused by old xmobars not being
terminated on early EOF, and is thus necessary on _all_ systems that
rely on EOF to terminate old xmobar before starting a new one. (To
trigger the bug, one additionally needs to close the xmobar pipe before
sending any input to it, which is unusual, but incorrectly configured
xmonad might trigger that.)
Furthermore, this fixes another execution path that could lead to xmobar
not being terminated on EOF:
`echo -e '\xff' | xmobar -c '[Run StdinReader]' -t '%StdinReader%'`
would terminate the StdinReader thread upon catching the "invalid
argument (invalid byte sequence)" so there'd be no thread to detect the
subsequent EOF and xmobar would get stuck.
Additionally, I believe that terminating either the thread or the entire
xmobar upon receiving a single miscoded byte isn't desirable, as this
might be an intermittent issue and another input line can be perfectly
okay. Therefore I suggest that the original issue @psibi was trying to
fix by b7a3d6745817 is worked around by introducing a throttling delay
instead of terminating the thread, as I assume that exceptions other
than async and EOF are recoverable.
Fixes: b7a3d6745817 ("Avoid busy looping by not catching all exceptions")
Fixes: 68ac4d3ae6f3 ("Update stderr and the bar on receiving exception")
Fixes: ed0663aac942 ("Add EOF check before getLine operation from stdin")
Fixes: https://github.com/jaor/xmobar/issues/442
Related: https://github.com/jaor/xmobar/pull/439
Related: https://github.com/jaor/xmobar/pull/448
|
|
This makes the Date plugin approximately twice as fast, and makes xmobar
up to about 5–10 % faster if Date is the only active plugin. (If more
expensive plugins like Network or MultiCpu are used, it doesn't make any
measurable difference.)
Micro-benchmark results on my HW:
Date Benchmarks/Date mean 2.833 μs ( +- 16.08 ns )
Date Benchmarks/DateZonedTime mean 5.020 μs ( +- 32.91 ns )
Date Benchmarks/DateWithTimeZone mean 2.827 μs ( +- 20.52 ns )
(DateZonedTime is the original implementation and DateWithTimeZone is
the implementation we had since 0.34 which never refreshes timezone.)
Real-life measurements (done overnight on an idle laptop, with all
measured xmobars running in parallel to ensure comparable conditions;
xmobars configured to only display date and with rate 10 — once per
second):
$ time timeout 6h xmobar .xmobarrc-DateZonedTime
real 360m0,010s
user 0m9,867s
sys 0m4,644s
(9.867 + 4.644) / (360 * 60) = 0.000672
$ time timeout 6h xmobar .xmobarrc-Date
real 360m0,008s
user 0m9,535s
sys 0m4,327s
(9.535 + 4.327) / (360 * 60) = 0.000642
$ time timeout 6h xmobar .xmobarrc-Date-10m
real 360m0,010s
user 0m9,780s
sys 0m4,215s
(9.780 + 4.215) / (360 * 60) = 0.000648
$ time timeout 6h xmobar .xmobarrc-DateWithTimeZone
real 360m0,006s
user 0m9,658s
sys 0m4,166s
(9.658 + 4.166) / (360 * 60) = 0.000640
(.xmobarrc-Date-10m is the proposed implementation, but with timezone
refresh every 10 minutes instead of every 1 minute)
Interpretation of these results:
* refreshing xmobar with just date takes around 650 μs
* that is xmobar with just date uses around 0.065 % of CPU time
* refreshing timezone takes additional cca 30 μs
When we only refresh timezone once a minute, these 30 μs become 0.5 μs
amortized, and that should be acceptable to even the most dedicated
perfectionist :-)
Fixes: a58e32f7c8af ("Revert "Optimize date plugin"")
Fixes: 878db3908060 ("Optimize date plugin")
Co-authored-by: Sibi Prabakaran <sibi@psibi.in>
|
|
This reverts commit 878db39080607ba476ba8d8f547ad28259efb6a9. That
commit optimized the date plugin by avoiding calling getTimeZone on each
execution, instead calling it just once upon startup and reusing that
zone. As a result, the time zone will not be updated dynamically, e.g.
when shifting in or out of daylight savings time.
I noticed this after my local time zone had shifted from EDT to EST. My
xmobar showed 4:30 when the local time was in fact 3:30 (and running
date on the command line confirmed that my system clock was actually
aware of this shift). I had to restart xmobar in order to pick up the
new time zone.
I repro'd the unexpected behavior by temporarily disabling my system's
time syncing, setting the time to 30 seconds before the zone shift,
running date to confirm I'd set the correct time, restarting xmobar, and
observing.
sudo systemctl stop systemd-timesyncd.service
date --set="01:59:30"
date
I observed my xmobar clock go from 1:59 to 2:00, rather than from 1:59
to 1:00 as expected.
Following the same steps, I was able to verify that this commit fixes
the issue.
|
|
spotifyd is funky that way
|
|
|
|
Both functions had a default parameter for use in some error cases. Now each accepts only one parameter (a
PerChannel), and return Nothing on an error.
The definition of 'channel' confused me, so I simplified it. Hopefully it's now
more clear that it just applies 'toInteger' to the 'IO (Maybe
CLong)' that 'channel'' returns.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
As documented in the http-client library, calling newManager is an
expensive operation:
```
Creating a new Manager is a relatively expensive operation, you are
advised to share a single Manager between requests instead.
```
But inspite of the haddocks in xmobar claiming that once 'Manager' is
created, it will be used throughout the monitor is not true. Because for
every call of `startWeather` a new manager is being created.
Also I removed the option in WeatherOpts because even if it is false,
it will be ultimately created in `getData` function. Also without
using a manager - the plugin won't really work. So, I don't think
there is any reason for this option to exist.
I have introduced a new dependency http-client-tls to use the shared
global manager so that we reuse the same manager every time. This
simplifies a lot of code. Note that this is not really a new
dependency because http-conduit already depends on it transitively.
|
|
We avoid calling getTimeZone for each of the time the date has to be
updated. Instead, it's computed once at the start and re-used for each
invocation.
Looking at the implementation of 'getTimeZone', we can see that it's
very expensive:
https://www.stackage.org/haddock/lts-15.15/time-1.9.3/src/Data-Time-LocalTime-Internal-TimeZone.html#getTimeZone
It calls a C FFI each time to get the time
zone (getTimeZoneCTime). This is something which we can avoid and the
MR implements that.
I have been using my xmobar with this patch and the result has been
quite good. My xmobar CPU usage has used to hit 3~7%
intermittently. With this MR, It hits only 0.7% intermittently which
is nice. :-)
|
|
|
|
Fixes https://github.com/jaor/xmobar/issues/442
|
|
|
|
|
|
|
|
|
|
This specifically avoids situation described in this issue
https://github.com/jaor/xmobar/issues/438 where the handle was
throwing the IOException continously in a loop:
<stdin>: hGetLine: invalid argument (invalid byte sequence)
It happened because my system's environment was right, but the proper
behaviour hear would be to let it to throw the exception rather than
leading to a busy loop.
I did some git blame to find out that this commit introduced the
behaviour:
https://github.com/jaor/xmobar/commit/fc24dc1874dcf7c9e66e21502a58b40cbe627c85
but there was no reason mentioned in the commit for trying to capture
all exceptions.
|
|
|
|
hGetLineSafe is always hGetLine and hence we can directly use it.
|
|
|
|
|
|
This adds a new `HandleReader` plugin, which displays data from a
Haskell `Handle`. This is really only useful if you are running xmobar
from within another Haskell program, but lets you avoid the mechanics of
creating a named pipe with the proper file permissions.
Instead, you can use `System.Process.createPipe` to make a pair of read
& write Handles. If you pass the read handle to HandleReader, you can
use hPutStr on the write Handle to send data to xmobar from your
application code.
|
|
|
|
This makes the code hlint-clean for --cpp-define=USE_NL80211,
--cpp-define=IWLIB and without --cpp-define too.
|
|
|
|
NL80211 was introduced in Linux 2.6.24 in 2007 as a new extensible
universal API, replacing "wireless extensions" ioctls. It works on top
of netlink, and allows direct communication to cfg80211 kernel
subsystem. Since then it became a hard requirement for all upstream
wireless drivers to hook into cfg80211 (SoftMAC drivers do it via the
common mac80211 layer). There's still additional compatibility code that
allows limited Wext functionality for cfg80211 drivers but it's buggy
and can be disabled altogether when CONFIG_CFG80211_WEXT is not set.
This patch makes use of "netlink" Haskell library which doesn't have any
additional runtime dependencies (so neither iwlib nor libnl are
required). The operation is the same as performed by "iw dev <devname>
link" command.
The signal level is transformed to "quality" by first clamping it to
[-110; -40], then adding 110 and dividing by 70 (same meaningless
formula as used by the cfg80211 Wext compatibility layer).
"essid" template argument is replaced by more appropriate "ssid" (with
the old variant still available for backwards compatibility)
|
|
A preparation for timer coalescing: tenthSeconds is just a sleep whereas
doEveryTenthSeconds enables using a central timer and waiting for all
monitors to update before refreshing the window. This commit is just a
simple refactor, the actual timer coalescing code comes later.
|
|
The calculations are based on the assumption that current_now (or
power_now) are always positive. However, power_supply documentation in
the kernel sources say nothing about it, and so some drivers provide a
signed value (e.g. bq27xxx_battery_current).
Discovered and fixed on ac100/paz00 netbook.
|
|
|
|
|
|
|
