Commit Graph
9 Commits
Author SHA1 Message Date
Chris DownandHiltjo Posthuma 712d6639ff Use sigaction(SA_NOCLDWAIT) for SIGCHLD handling
signal() semantics are pretty unclearly specified. For example, depending on OS
kernel and libc, the handler may be returned to SIG_DFL (hence the inner call
to read the signal handler). Moving to sigaction() means the behaviour is
consistently defined.

Using SA_NOCLDWAIT also allows us to avoid calling the non-reentrant function
die() in the handler.

Some addditional notes for archival purposes:

* NRK pointed out errno of waitpid could also theoretically get clobbered.
* The original patch was iterated on and modified by NRK and Hiltjo:
  * SIG_DFL was changed to SIG_IGN, this is required, atleast on older systems
    such as tested on Slackware 11.
  * signals are not blocked using sigprocmask, because in theory it would
    briefly for example also ignore a SIGTERM signal. It is OK if waitpid() is (in
    theory interrupted).

POSIX reference:
"Consequences of Process Termination":
https://pubs.opengroup.org/onlinepubs/9699919799/functions/_Exit.html#tag_16_01_03_01
2023-01-28 13:34:43 +01:00
Chris DownandHiltjo Posthuma 89f9905714 grabkeys: Avoid missing events when a keysym maps to multiple keycodes
It's not uncommon for one keysym to map to multiple keycodes. For
example, the "play" button on my keyboard sends keycode 172, but my
bluetooth headphones send keycode 208, both of which map back to
XF86AudioPlay:

    % xmodmap -pke | grep XF86AudioPlay
    keycode 172 = XF86AudioPlay XF86AudioPause XF86AudioPlay XF86AudioPause
    keycode 208 = XF86AudioPlay NoSymbol XF86AudioPlay
    keycode 215 = XF86AudioPlay NoSymbol XF86AudioPlay

This is a problem because the current code only grabs a single one of
these keycodes, which means that events for any other keycode also
mapping to the bound keysym will not be handled by dwm. In my case, this
means that binding XF86AudioPlay does the right thing and correctly
handles my keyboard's keys, but does nothing on my headphones. I'm not
the only person affected by this, there are other reports[0].

In order to fix this, we look at the mappings between keycodes and
keysyms at grabkeys() time and pick out all matching keycodes rather
than just the first one. The keypress() side of this doesn't need any
changes because the keycode gets converted back to a canonical keysym
before any action is taken.

0: https://github.com/cdown/dwm/issues/11
2022-12-07 23:06:26 +01:00
Chris DownandHiltjo Posthuma 8b48e30973 manage: Make sure c->isfixed is applied before floating checks
Commit 8806b6e237 ("manage: propertynotify: Reduce cost of unused size
hints") mistakenly removed an early size hints update that's needed to
populate c->isfixed for floating checks at manage() time. This resulted
in fixed (size hint min dimensions == max dimensions) subset of windows
not floating when they should.

See https://lists.suckless.org/dev/2204/34730.html for discussion.
2022-04-26 15:50:55 +02:00
Chris DownandHiltjo Posthuma 8806b6e237 manage: propertynotify: Reduce cost of unused size hints
This patch defers all size hint calculations until they are actually
needed, drastically reducing the number of calls to updatesizehints(),
which can be expensive when called repeatedly (as it currently is during
resizes).

In my unscientific testing this reduces calls to updatesizehints() by
over 90% during a typical work session. There are no functional changes
for users other than an increase in responsiveness after resizes and
a reduction in CPU time.

In slower environments or X servers, this patch also offers an
improvement in responsiveness that is often tangible after resizing a
client that changes hints during resizes.

There are two main motivations to defer this work to the time of hint
application:

1. Some clients, especially terminals using incremental size hints,
   resend XA_WM_NORMAL_HINTS events on resize to avoid fighting with the
   WM or mouse resizing. For example, some terminals like urxvt clear
   PBaseSize and PResizeInc during XResizeWindow and restore them
   afterwards.

   For this reason, after the resize is concluded, we typically receive
   a backlogged XA_WM_NORMAL_HINTS message for each update period with
   movement, which is useless. In some cases one may get hundreds or
   thousands of XA_WM_NORMAL_HINTS messages on large resizes, and
   currently all of these result in a separate updatesizehints() call,
   of which all but the final one are immediately outdated.

   (We can't just blindly discard these messages during resizes like we
   do for EnterNotify, because some of them might actually be for other
   windows, and may not be XA_WM_NORMAL_HINTS events.)

2. For users which use resizehints=0 most of these updates are unused
   anyway -- in the normal case where the client is not floating these
   values won't be used, so there's no need to calculate them up front.

A synthetic test using the mouse to resize a floating terminal window
from roughly 256x256 to 1024x1024 and back again shows that the number
of calls to updatesizehints() goes from over 500 before this patch (one
for each update interval with movement) to 2 after this patch (one for
each hint application), with no change in user visible behaviour.

This also reduces the delay before dwm is ready to process new events
again after a large resize on such a client, as it avoids the thundering
herd of updatesizehints() calls when hundreds of backlogged
XA_WM_NORMAL_HINTS messages appear at once after a resize is finished.
2022-04-16 16:37:46 +02:00
Chris DownandHiltjo Posthuma 8657affa2a drawbar: Don't expend effort drawing bar if it is occluded
I noticed that a non-trivial amount of dwm's work on my machine was from
drw_text, which seemed weird, because I have the bar disabled and we
only use drw_text as part of bar drawing.

Looking more closely, I realised that while we use m->showbar when
updating the monitor bar margins, but don't skip actually drawing the
bar if it is hidden. This patch skips drawing it entirely if that is the
case.

On my machine, this takes 10% of dwm's on-CPU time, primarily from
restack() and focus().

When the bar is toggled on again, the X server will generate an Expose
event, and we'll redraw the bar as normal as part of expose().
2021-12-19 16:16:30 +01:00
Chris DownandHiltjo Posthuma 67d76bdc68 Do not allow focus to drift from fullscreen client via focusstack()
It generally doesn't make much sense to allow focusstack() to navigate
away from the selected fullscreen client, as you can't even see which
client you're selecting behind it.

I have had this up for a while on the wiki as a separate patch[0], but
it seems reasonable to avoid this behaviour in dwm mainline, since I'm
struggling to think of any reason to navigate away from a fullscreen
client other than a mistake.

0: https://dwm.suckless.org/patches/alwaysfullscreen/
2021-03-29 19:16:27 +02:00
Chris DownandHiltjo Posthuma ed3ab6b4fc drawbar: Don't shadow sw global
This jarred me a bit while reading the code, since "sw" usually refers
to the global screen geometry, but in drawbar() only it refers to
text-related geometry. Renaming it makes it more obvious that these are
not related.
2020-04-22 20:33:39 +02:00
Chris DownandHiltjo Posthuma f087d20e6e getatomprop: Add forward declaration
No functional changes, but for every other function we have a forward
declaration here. getatomprop should be no exception.
2020-04-22 20:33:26 +02:00
Chris DownandHiltjo Posthuma a8e9513783 setmfact: Unify bounds for compile-time and runtime mfact
There are two places that mfact can be set:

- In the mfact global, which is defined at compile time and passed
  into m->mfact during monitor setup. No bounds checks are performed,
  but the comment alongside it says that valid values are [0.05..0.95]:

      static const float mfact = 0.55; /* factor of master area size [0.05..0.95] */

- By setmfact, which adjusts m->mfact at runtime. It also does some
  minimum and maximum bounds checks, allowing [0.1..0.9]. Values outside
  of that range are ignored, and mfact is not adjusted.

These different thresholds mean that one cannot setmfact 0.95 or 0.05,
despite the comment above that lists the legal range for mfact.

Clarify this by enforcing the same bounds in setmfact at runtime as
those listed for mfact at compile time.
2020-04-20 17:56:41 +02:00