itself hit the "terminal can't keep up" check. To avoid this, record how
much data we send during redraw (we know we will be starting with 0) and
skip the check until it has been flushed. GitHub issue 912.
xterm-keys by default, generates \033[1;3A instead of
\033\033[OA. Unfortunately this confuses vi, which doesn't understand
xterm keys and now sees Escape+Up pressed within escape-time as Escape
followed by A.
The issue doesn't happen in xterm itself because it gets the keys from X
and can distinguish between a genuine M-Up and Escape+Up.
Because xterm can, tmux can too: xterm will give us \033[1;3A (that is,
kUP3) for a real M-Up and \033\033OA for Escape+Up - in fact, we can be
sure any \033 preceding an xterm key is a real Escape key press because
Meta would be part of the xterm key instead of a separate \033.
So change tmux to recognise both sequences as M-Up for its own purposes,
but generate the xterm version of M-Up only if it originally received
the xterm version from the terminal.
This means we will return to sending \033\033OA instead of the xterm key
for terminals that do not support xterm keys themselves, but there is no
practical way around this because they do not allow us to distinguish
between Escape+Up and M-Up. xterm style escape sequences are now the de
facto standard for these keys in any case.
Problem reported by jsing@ and subsequently by Cecile Tonglet in GitHub
issue 907.
and not have to wait for an update when they change pane, we allow
commands to run more than once a second if the expanded form
changes. Unfortunately this can mean them being run far too often
(pretty much continually) when multiple clients exist, because some
formats (including #D) will always differ between clients.
To avoid this, give each client its own tree of jobs which means that
the same command will be different instances for each client - similar
to how we have the tag to separate commands for different panes.
GitHub issue 889; test case reported by Paul Johnson.
command. This is used for the session, window and pane for all commands
in the command sequence if there is no -t or -s.
However, using it for all commands in the command sequence means that if
the active pane or current session is changed, subsequent commands still
use the previous state. So make commands which explicitly change the
current state (such as neww and selectp) update it themselves for later
commands. Commands which may invalidate the state (like killp) are
already OK because an invalid state will be ignored.
Also fill in the current state for all key bindings rather than just the
mouse, so that any omissions are easier to spot.
will suppress root key table bindings. So change to always check the
root table if no binding is found in the current table (whether it be
the prefix table from pressing the prefix or the copy mode table from a
pane).
A root key binding can be blocked by binding the key to a command that
does nothing (like send-keys with no arguments).
Problem reported by Thomas Sattler.
bind q select-pane -U \; resize-pane -Z
(There is still some possible weirdness with the way we do current
targets, it should probably be done in a different way at some point.)
reasonable amount (currently width * height * 8 bytes), discard all
output to the terminal and start trying to redraw periodically
instead. Continue with this until the amount of data we are trying to
write falls to a low level again.
This helps to prevent tmux sitting on a huge buffer of data when there
are processes with fast output running inside tmux but the outside
terminal is slow.
A new client_discarded format holds the amount of data that has been
discarded due to this mechanism.
The three variables (when to start this, when to stop, and how often to
redraw) are basically "works for me" at the moment, this is going in to
see how it goes and if it causes problems for anyone else.
with of the terminal scrolls, tmux needs to redraw the entire pane. This
results in a large amount of output data which can cause slow terminals
to struggle, particularly when many lines are scrolled together quickly.
This can be reduced by only redrawing when tmux doesn't hold any
buffered data for the terminal. If a redraw is required and data is
buffered, the redraw is deferred until all that data is consumed (it is
checked after every event loop, a timer is used to ensure this happens
at some point). While a redraw is pending, no additional data will be
written to the terminal.
The redraw still happens, now it is just pushed back if it is possible
it would just add more data on top of a terminal that is already
behind. This both gives the terminal a chance to catch up, and allows
tmux to process more scrolling (that would require additional redraws)
in the meantime.
Helps with a problem reported by Greg Hurrell.
falls back to an alternative if the tty name is not available. This is
clearer than overloading the client ttyname member and allows us to
remove the path stored in the tty struct, it should always be the same
as the client.
jobs, this means that if the same job is used for different windows or
panes (for example in pane-border-format), it will be run separately for
each pane.
main issue is that if we have two panes, A with 1002 and B with 1003, we
need to set 1003 outside tmux in order to get all the mouse events, but
then we need to suppress the ones that pane A doesn't want. This is easy
in SGR mouse mode, because buttons == 3 is only used for movement events
(for other events the trailing m/M marks a release instead), but in
normal mouse mode we can't tell so easily. So for that, look at the
previous event instead - if it is drag+release as well, then the current
event is a movement event.
so too early, before the mouse event was necessarily valid, so could end
up using the pane from the previous mouse event, or the active pane.
It is important that we use the right pane now that different panes can
have different key tables (for copy mode).
Fixes problem reported by Greg Hurrell.
supports them (that is, if it advertises itself as a VT420 - probably
just xterm). These are the vertical equivalent of the scroll region and
allow much faster scrolling of panes that do not take up the full width
of the terminal.
add a link of winlinks to each window and a pointer to the session to
each winlink. Also rewrite the alerts processing to return to the old
behaviour (alert in any window sets the flag on any winlink).
but there is also now a global command queue. Instead of command queues
being dispatched on demand from wherever the command happens to be
added, they are now all dispatched from the top level server
loop. Command queues may now also include callbacks as well as commands,
and items may be inserted after the current command as well as at the end.
This all makes command queues significantly more predictable and easier
to use, and avoids the complex multiple nested command queues used by
source-file, if-shell and friends.
A mass rename of struct cmdq to a better name (cmdq_item probably) is
coming.
support line editing and instead executes a command as soon as a
non-number key is pressed. Add a -N flag to command-prompt for the same
in copy mode. Reported by Theo Buehler.
The vi-copy and emacs-copy mode key tables are gone, and instead copy
mode commands are bound in one of two normal key tables ("copy-mode" or
"copy-mode-vi"). Keys are bound to "send-keys -X copy-mode-command". So:
bind -temacs-copy C-Up scroll-up
bind -temacs-copy -R5 WheelUpPane scroll-up
Becomes:
bind -Tcopy-mode C-Up send -X scroll-up
bind -Tcopy-mode WheelUpPane send -N5 -X scroll-up
This allows the full command parser and command set to be used - for
example, we can use the normal command prompt for searching, jumping,
and so on instead of a custom one:
bind -Tcopy-mode C-r command-prompt -p'search up' "send -X search-backward '%%'"
command-prompt also gets a -1 option to only require on key press, which
is needed for jumping.
The plan is to get rid of mode keys entirely, so more to come eventually.
pane-border-status is set to "top" or "bottom" (rather than "off"),
every pane has a permanent top or bottom border containing the text from
pane-border-format.
Based on a diff sent long ago by Jonathan Slenders, mostly rewritten and
simplified by me.
poor idea that was fairly quickly replaced by SGR mouse input (which is
now widespread). It is impossible to tell the difference between UTF-8
and non-UTF-8 mouse input; since the mouse-utf8 option was removed tmux
has not handled it correctly in any case; and it is ridiculous to have
three different forms of mouse input.
uint64_t and converting UTF-8 to Unicode on input and the reverse on
output. (This allows key bindings, there are still omissions - the
largest being that the various prompts do not accept UTF-8.)
around, we can't use file descriptors for the working directory because
we will be unable to pass it to a privileged process to tell it where to
read or write files or spawn children. So move tmux back to using
strings for the current working directory. We try to check it exists
with access() when it is set but ultimately fall back to ~ if it fails
at time of use (or / if that fails too).
use event_once to queue a callback to deal with them. Also dead clients
with references would never actually be freed because the wrap-up
functions (the callback for stdin, or status_prompt_clear) would never
be called. So call them in server_client_lost.
keys. The default key bindings become the "prefix" table and -n the
"root" table. Keys may be bound in new tables with bind -T and
switch-client -T used to specify the table in which the next key should
be looked up. Based on a diff from Keith Amling.
options for "mouse-this" and "mouse-that", mouse events may be bound as
keys and there is one option "mouse" that turns on mouse support
entirely (set -g mouse on).
See the new MOUSE SUPPORT section of the man page for description of the
key names and new flags (-t= to specify the pane or window under mouse
as a target, and send-keys -M to pass through a mouse event).
The default builtin bindings for the mouse are:
bind -n MouseDown1Pane select-pane -t=; send-keys -M
bind -n MouseDown1Status select-window -t=
bind -n MouseDrag1Pane copy-mode -M
bind -n MouseDrag1Border resize-pane -M
To get the effect of turning mode-mouse off, do:
unbind -n MouseDrag1Pane
unbind -temacs-copy MouseDrag1Pane
The old mouse options are now gone, set-option -q may be used to
suppress warnings if mixing configuration files.
while and in fact it is less useful that using the client ttyname. So
don't bother and don't pass it from the client. If we need it in future
it is in c->environ.
descriptors rather than strings.
- Each session still has a current working directory.
- New sessions still get their working directory from the client that
created them or its attached session if any.
- New windows are created by default in the session working directory.
- The -c flag to new, neww, splitw allows the working directory to be
overridden.
- The -c flag to attach let's the session working directory be changed.
- The default-path option has been removed.
To get the equivalent to default-path '.', do:
bind c neww -c $PWD
To get the equivalent of default-path '~', do:
bind c neww -c ~
This also changes the client identify protocol to be a set of messages rather
than one as well as some other changes that should make it easier to make
backwards-compatible protocol changes in future.
window or unzoom (restored to the normal layout) if it already zoomed,
bound to C-b z by default. The pane is unzoomed on pretty much any
excuse whatsoever.
We considered making this a new layout but the requirements are quite
different from layouts so decided it is better as a special case. Each
current layout cell is saved, a temporary one-cell layout generated and
all except the active pane set to NULL.
Prompted by suggestions and scripts from several. Thanks to Aaron Jensen
and Thiago Padilha for testing an earlier version.
commands and allow a command to block execution of subsequent
commands. This allows run-shell and if-shell to be synchronous which has
been much requested.
Each client has a default command queue and commands are consumed one at
a time from it. A command may suspend execution from the queue by
returning CMD_RETURN_WAIT and then resume it by calling cmd_continue() -
for example run-shell does this from the callback that is fired after
the job is freed.
When the command queue becomes empty, command clients are automatically
exited (unless attaching). A callback is also fired - this is used for
nested commands in, for example, if-shell which can block execution of
the client's cmdq until a new cmdq becomes empty.
Also merge all the old error/info/print functions together and lose the
old curclient/cmdclient distinction - a cmdq is bound to one client (or
none if in the configuration file), this is a command client if
c->session is NULL otherwise an attached client.
add a new value to mean "leave client running but don't attach" to fix
problems with using some commands in a command sequence. Most of the
work by Thomas Adam, problem reported by "jspenguin" on SF bug 3535531.
commands be sent and output received on stdout. This can be used to
integrate with other terminal emulators and should allow some other
things to be made simpler later. More to come so doesn't do much yet and
deliberately not documented.
handling them in the server, handle them in the client and pass buffers
over imsg. This is much tidier for some upcoming changes and the
performance hit isn't critical.
The tty fd is still passed to the server as before.
This bumps the tmux protocol version so new clients and old servers are
incompatible.
separate options, prefix and prefix2. This simplifies the code and gets
rid the data options type which was only used for this one option.
Also add a -2 flag to send-prefix to send the secondary prefix key,
fixing a cause of minor irritation.
People who want three prefix keys are out of luck :-).
be a huge rush of it after they are resumed/unlocked. The main output
path was fine but status line updates and the terminal state reset code
were missed.
mode when the mouse is dragged or the mouse wheel is used. Also exit
copy mode when the mouse wheel is scrolled off the bottom. Discussed
with and written by hsim at gmx dot li.
fire-and-forget.
Status jobs now managed with two trees of output (new and old), rather
than storing the output in the jobs themselves. When the status line is
processed any jobs which don't appear in the new tree are started and
the output from the old tree displayed. When a job finishes it updates
the new tree with its output and that is used for any subsequent
redraws. When the status interval expires, the new tree is moved to the
old so that all jobs are run again.
This fixes the "#(echo %H:%M:%S)" problem which would lead to thousands
of identical persistent jobs and high memory use (this can still be
achieved by adding "sleep 30" but that is much less likely to happen by
accident).
and supports larger terminals than the older way.
If the new mouse-utf8 option is on, UTF-8 mouse input is enabled for all
UTF-8 terminals. The option defaults to on if LANG etc are set in the
same manner as the utf8 option.
With help and based on code from hsim at gmx.li.
this screws up the choice of most-recently-used. Instead, break the time
update into a little function and do it when the session is attached.
Pointed out by joshe@.
- server option "exit-unattached" makes the server exit when no clients
are attached, even if sessions are present;
- session option "destroy-unattached" destroys a session once no clients
are attached to it.
These are useful for preventing tmux remaining in the background where
it is undesirable and when using tmux as a login shell to keep a limit
on new sessions.
when the client tty backs up too much, just stop updating the tty and
only update the internal screen. Then when the tty recovers, force a
redraw.
This prevents a dodgy client from causing other clients to go into
backoff while still allowing tmux to be responsive (locally) when seeing
lots of output.
stdin up to the server, I forgot one essential point - the tmux server
could now be both the producer and consumer. This happens when tmux is
run inside tmux, as well as when piping tmux commands together.
So, using stdio(3) was a bad idea - if sufficient data was written, this
could block in write(2). When that happened and the server was both
producer and consumer, it deadlocks.
Change to use libevent bufferevents for the client stdin, stdout and
stderr instead. This is trivial enough for output but requires a
callback mechanism to trigger when stdin is finished.
This relies on the underlying polling mechanism for libevent to work
with whatever devices to which the user could redirect stdin, stdout or
stderr, hence the change to use poll(2) over kqueue(2) for tmux.
commands can directly make use of them. This means that load-buffer and
save-buffer can have "-" as the file to read from stdin or write to stdout.
This is a protocol version bump so the tmux server will need to be restarted
after upgrade (or an older client used).
starting tmux from .xinitrc.
One of the very few things the server relies on the client for now is to
pass through a message on SIGWINCH, but there is a condition where
potentially a SIGWINCH may be lost during the transition from unattached
(main.c) to attached (client.c). So trigger a size change immediately
after the client installs its SIGWINCH handler.
Also, when the terminal is resized, reset the scroll region and cursor
position. Previously, we were clearing our saved idea of these, but in
fact some terminals do not reset them on resize, so this caused problems
during redraw.
While here make a resize to the same size not cause a redraw and rename
the tmux.out output log file to include the tmux PID.
function. We were only ever using the client to find the session anyway.
This allows send-key to work properly for manipulating copy mode from
outside tmux.
From Micah Cowan.
