Modern Linux kernels send additional flags (FATTR_CTIME, etc.) along
with FATTR_SIZE when truncating files. The exact equality check
(in->valid == FATTR_SIZE) failed, causing shell redirections like
'echo foo > file' to return ENOSYS on existing files.
Change to a bitmask check: (in->valid & FATTR_SIZE)
This indicates window focus change. Without this, focus change caused by alt-key combinations (ctrl+alt+arrow on many Linux configurations will switch virtual desktops) may incorrectly be interpreted as a key-compose.
There has been some funny locking/stalling behavior that seems to be related to using alt+click as a simulated middle-click. It may have started after this gfx_abortcompose call was added. According to devdraw.h, the gfx_ routines should not be called with a lock held. Now I'm not sure this fixes the weird behavior I've seen, but can't hurt to still fix this buggy call.
This is OK with the touch pad, but it's actually not great with a real mouse scroll wheel, because the scrolling becomes inconsistent and hard to predict exactly how much the screen will scroll. Better to just remove it.
The alt key begins a compose. When using alt+click, compose was not previously cancelled and the next two typed characters would often end up composed.
Upgrade to version 4 of the wl_seat interface which adds the
wl_keyboard::repeat_info event allowing the compositor to configure
key repeat delay and rate.
This change splits scroll repeat and key repeat into separate
callbacks, and only registers them when a key is pressed or
scrolling is active, respectively. The key repeat logic is
changed to support key repetitions higher than the frame rate.
The new logic will also accomodate key repeat parameters set by
the compositor, although the comment about us not receiving that
configuration event is still valid.
I found sway 1.10 and wlroots 0.18.2 will occasionally generate
wl_keyboard::keymap events with an empty keymap. I do not have
a reliable reproduction for this, but it seems to occur when
switching back and forth between workspaces with the keyboard.
This also fixes the issue where devdraw can go into a busy loop if
the compositor goes away and wl_display_dispatch() returns -1.
I tried running the wayland devdraw implementation on Sway 1.10,
wlroots 0.18 and encountered the error:
xdg_surface#13: error 3: xdg_surface has never been configured
According to https://wayland.app/protocols/xdg-shell#xdg_surface ,
a client must commit a surface without a buffer, *wait* for the first
configure request from the compositor, ack it, and *only then* can
it proceed to attach a buffer to the surface and tell wayland to display it.
This was caused by the following sequence of events:
1. devdraw starts, enters gfx_main
2. `gfx_main` calls `gfx_started`, which spawns the `serveproc` thread
3. `gfx_main` enters `wl_display_dispatch`, flushing any buffered requests to
the compositor, and enters `wl_display_poll()` to wait for incoming messages
4. `serveproc` calls `rpc_attach`, sets up the surface, and buffers a commit.
The race is between #3 and #4. If #3 happens first, the buffered commit
just sits there until `rpc_flush` is called, which calls `wl_display_flush()`,
but at that point a buffer is attached too quickly for the configure to happen.
This commit fixes the race by adding a `configured` field to the WaylandClient
and using it to guard `rpc_flush`.
In addition, I found that mouse warping, at least in sway, would move
the cursor but future mouse presses would register at the old location
until I moved the mouse. So I added a call to gfx_mousetrack to the end
of `rpc_setmouse`.
I uncovered this problem while writing unit tests for GNU troff's
delimited expression handling. Plan 9 troff's numeric expression
evaluator handles division by zero but not modulus by zero.
Fixes:
$ echo '.if %0%0% .tm true' | 9 troff
Floating point exception (core dumped)
$ echo '.if 1%0 .tm true' | 9 troff
Floating point exception (core dumped)
After this patch:
$ echo '.if %0%0% .tm true' | 9 troff
x T utf
x res 720 1 1
x init
troff: modulus by zero.; stdin:1
troff: modulus by zero.; stdin:1
x trailer
V0
x stop
$ echo '.if 1%0 .tm true' | 9 troff
x T utf
x res 720 1 1
x init
troff: modulus by zero.; stdin:1
x trailer
V0
x stop
x11-screen.c accepts selection targets with the MIME type
"text/plain;charset-UTF8", but it should do so case insensitively. The
current code draws an error when working with xfce4-terminal 1.0.4,
which sends lower case. To reproduce, run acme, snarf some text, then
paste in the terminal. acme will report:
acme: cannot handle selection request for 'text/plain;charset=utf-8' (590)
...when rendering some man pages, such as those of ncurses.
I did not manage (nor seriously attempt) to identify the root cause of
this bug. ncurses's use of `SH` and `SS` man(7) macros is
unremarkable.[1] I cannot account for why the less(1) man page renders
fine and ncurses pages like insstr(3) do not. But render badly they do,
emitting *roff logic as formatted output.
```
$ 9 nroff -man $(man -w insstr) | grep -F .ss | cat -v
"'''if^GNAME^GSYNOPSIS^G .ss 18 NAME
"'''if^GSYNOPSIS^GSYNOPSIS^G .ss 18 SYNOPSIS
"'''if^GDESCRIPTION^GSYNOPSIS^G .ss 18 DESCRIPTION
"'''if^GRETURN^GSYNOPSIS^G .ss 18 RETURN VALUE
"'''if^GNOTES^GSYNOPSIS^G .ss 18 NOTES
"'''if^GEXTENSIONS^GSYNOPSIS^G .ss 18 EXTENSIONS
"'''if^GPORTABILITY^GSYNOPSIS^G .ss 18 PORTABILITY
"'''if^GHISTORY^GSYNOPSIS^G .ss 18 HISTORY
"'''if^GSEE^GSYNOPSIS^G .ss 18 SEE ALSO
```
With this patch:
```
$ 9 nroff -man $(man -w insstr) | grep -F .ss | cat -v | grep . || echo NO OUTPUT
NO OUTPUT
```
I do observe that the problem seems to correspond to the only use in the
package of the old-fashioned `'''` commenting convention _within a macro
definition_. I have a notion of how GNU troff works, but little about
AT&T troff. That said, if I were to try to get to the bottom of this
problem, I'd look into if and how the no-break command character is
handled differently in copy mode. I see nothing in CSTR #54 to suggest
that the command characters have different meanings in copy mode and its
complement.[2]
My solution is to use idiomatic comment syntax inside macro definitions.
[1] d5dc8a4a7c/man/curs_insstr.3x (L47)
[2] unnamed in AT&T documentation but which I term "interpretation mode"
in groff
Because of a dangling else, after adding a name to an image with 'N',
it was immediately deleted, resulting in Enoname in response to 'n'.
This went mostly unnoticed, since plan9port itself does not use named
images, but it can break external applications like truedraw[1].
[1]: https://git.sr.ht/~arusekk/truedraw
Fixes: c66b52501b ("new draw server")
groff 1.23.0 added .MR to its -man macro package. The NEWS file states
that the inclusion of the macro "was prompted by its introduction to
Plan 9 from User Space's troff in August 2020." From d32deab it seems
that the name for Plan 9 from User Space's implementation was suggested
by groff maintainer G. Brandon Robinson.
Not sure if the intention was to make these definitions compatible, but
it would be nice if they were.
Currently, Plan 9 from User Space's .MR expects its second argument to
be parenthesized. groff's .MR does not. This results in extra
parentheses appearing in manual references when viewing Plan 9 from User
Space's manual pages on a system using groff.
Dependabot automatically maintains GitHub Actions workflows by sending
Pull Requests to project when action updates are released. Dependabot
labels the pull requests 'ci: bump actions/checkout from v3 to v4'.
Dependabot opens a maximum of five Pull Requests at a time and reviews
the need for PRs once a week.
Give `-Wformat=0` on illumos/Solaris so that the build does not
complain about `execlp` not having a sentinel (it's nil). Also
fix an integer size issue when casting to a void pointer.
libthread had code in `_pthreadinit` to test for LinuxThreads,
and error if it was in use. This hasn't been relevant since
Linux 2.6, and I can't imagine there are many people trying to
build (recent) plan9port on systems that old.
Unfortunately, this code violated aliasing rules, and compilers
were complaining about it. But, since it likely hasn't had much
relevance in 20ish years, we can probably just remove it.
Use `extern` where appropriate so that we can remove `-fcommon`
on macOS and others. On macOS, `-fcommon` generates a number of
linker warnings.
I've tested building on macOS, Linux, FreeBSD, NetBSD, OpenBSD
and illumos. I am not in a position to test on AIX or other
more esoteric platforms, however.
Misleading indentation warnings are actually useful, as they
will occasionally reveal actual bugs (cf the apple `goto fail`
security bug from a few years ago). Newer versions of clang
(at least) are more aggressive about warnings in this regard,
which has exposed a few warnings that are annoying. However,
they are easy to address.
Set defaults for `awk` and `egrep`, instead of
setting things multiple times when specializing on
OS, and using a default in a `case` statement.
This is cleaner and less error-prone.
Mostly turning the crank on fixing simple warnings: arrays, for
instance, can never be nil. A couple of pointers should have been
initialized to `nil` before being tested.
Some logic in `troff` was simplified: basically, an `if` statement had
a condition that must have always been true if that section of code
were being executed at all.
Introduce a new variable set in $PLAN9/config:
$egrep, which is conditionally set to either
`egrep` or `grep -E`, based on platform, and use
this when we want to invoke `egrep`.
This works around the endless warnings from GNU
grep stating that `egrep` is deprecated in favor
of `grep -E`, when run on e.g. Linux.
This bug is an intermittent annoyance and does not seem likely to be
fixed soon. It means that every so often my acme becomes unusable
because it behaves as if a mouse button is permanently set.
Avoid that issue by keeping track of mouse button state directly.
While we're about it, fix another annoying issue: under Linux the
FocusOut event reset the kstate field but not the kbuttons field,
meaning that the buttons state could inappropriately persist. This
happened occasionally by accident, but was easy to reproduce: with the
mouse button held down, drag outside the window and click Alt-Tab to
switch away from the window. We see a KeyDown event followed by FocusOut
but no KeyUp event.
Change-Id: I3708316478c80c41806d7157ae9929c6cd84b662
Levi Neely
Ethan Burns
Levi Neely
David Arroyo
ylh
G. Branden Robinson
Scott Schwartz
Dan Cross
Russ Cox
Arusekk
Anthony Sorace
Jordan Niethe
dependabot[bot]
Dmitri Vereshchagin
Kare Nuorteva
lufia
erentar
Roger Peppe
Igor Burago