When changing so that the client passes its stdout and stderr as well as
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.
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.
Return the command client return code with MSG_EXIT now that MSG_ERROR and
MSG_PRINT are unused.
New clients should be compatible with old tmux servers but vice versa may print
an error.
Custom layouts. list-windows command displays the layout as a string (such as
"bb62,159x48,0,0{79x48,0,0,79x48,80,0}") and it can be applied to another
window (with the same number of panes or fewer) using select-layout.
Send all three of stdin, stdout, stderr from the client to the server, so that
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).
Store the current working directory in the session, change the default-path
option to default to empty and make that mean that the stored session CWD is
used.
Setting the cmdlist pointer in the bind-key to NULL to prevent it being freed
after the command is executing is bogus because it may still be needed if the
same command is going to be executed again (for example if you "bind-key a
bind-key b ..."). Making a copy is hard, so instead add a reference count to
the cmd_list.
While here, also print bind-key -n and the rest of the flags properly.
Fixes problem reported by mcbride@.
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).
after the command is executing is bogus because it may still be needed if the
same command is going to be executed again (for example if you "bind-key a
bind-key b ..."). Making a copy is hard, so instead add a reference count to
the cmd_list.
While here, also print bind-key -n and the rest of the flags properly.
Fixes problem reported by mcbride@.
Fix problems with window sizing seen by Raghavendra D Prabhu when
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.
When the mode-mouse option is on, support dragging to make a selection
in copy mode.
Also support the scroll wheel, although xterm strangely does not ignore
it in application mouse mode, causing redraw artifacts when scrolling up
(other terminals appear to be better behaved).
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.
in copy mode.
Also support the scroll wheel, although xterm strangely does not ignore
it in application mouse mode, causing redraw artifacts when scrolling up
(other terminals appear to be better behaved).
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.
Identical behaviour to select-prompt can now be obtained with
command-prompt, so remove select-prompt and change ' to be bound to
command-prompt -p index "select-window -t :%%".
Make signal handler setup/teardown two common functions instead of six,
and reset SIGCHLD after fork to fix problems with some shells. From
Romain Francoise.
Permit keys in copy mode to be prefixed by a repeat count, entered with
[1-9] in vi mode, or M-[1-9] in emacs mode.
From Micah Cowan, tweaked a little by me.
copy mode uses the real screen as backing and if it is updated while copying,
strange things can happen. So, freeze reading from the pty while in copy mode.
Instead of bailing out on the first configuration file error, carry on,
collecting all the errors, then start with the active window in more mode
displaying them.
vi-style B, W and E keys in copy mode to navigate between words treating only
spaces as word separators. Also add . to the list of word separators for
standard word navigation.
From Micah Cowan, tweaked slightly by me.
Add "server options" which are server-wide and not bound to a session or
window. Set and displayed with "set -s" and "show -s".
Currently the only option is "quiet" (like command-line -q, allowing it to be
set from .tmux.conf), but others will come along.
window. Set and displayed with "set -s" and "show -s".
Currently the only option is "quiet" (like command-line -q, allowing it to be
set from .tmux.conf), but others will come along.
Massive spaces->tabs and trailing whitespace cleanup, hopefully for the last
time now I've configured emacs to make them displayed in really annoying
colours...
Eliminate duplicate code and ease the passage for server-wide options by adding
a -w flag to set-option and show-options and making setw and showw aliases to
set -w and show -w.
Note: setw and showw are still there, but now aliases for set -w and show -w.
a -w flag to set-option and show-options and making setw and showw aliases to
set -w and show -w.
Note: setw and showw are still there, but now aliases for set -w and show -w.
Change status line drawing to create the window list in a separate screen and
then copy it into the status line screen. This allows UTF-8 in window names and
fixes some problems with #[] in window-status-format.
Two new options, window-status-format and window-status-current-format, which
allow the format of each window in the status line window list to be controlled
using similar # sequences as status-left/right.
This diff also moves part of the way towards UTF-8 support in window names but
it isn't quite there yet.
Cleanup by moving various (mostly horrible) little bits handling UTF-8 grid
data into functions in a new file, grid-utf8.c, and use sizeof intead of
UTF8_DATA.
Also nuke trailing whitespace from tmux.1, reminded by jmc.
Add a per-client log of status line messages displayed while that client
exists. A new message-limit session option sets the maximum number of entries
and a command, show-messages, shows the log (bound to ~ by default).
This (and prompt history) might be better as a single global log but until
there are global options it is easier for them to be per client.
allow the format of each window in the status line window list to be controlled
using similar # sequences as status-left/right.
This diff also moves part of the way towards UTF-8 support in window names but
it isn't quite there yet.
exists. A new message-limit session option sets the maximum number of entries
and a command, show-messages, shows the log (bound to ~ by default).
This (and prompt history) might be better as a single global log but until
there are global options it is easier for them to be per client.
Switch the tty key tree over to an (unbalanced) ternary tree which allows
partial matches to be done (they wait for further data or a timer to expire,
like a naked escape).
Mouse and xterm-style keys still expect to be atomic.
Switch tty key input over to happen on a read event. This is a bit more
complicated because of escape input, but in that case instead of processing a
key immediately, schedule a timer and reprocess the bufer when it expires.
This currently assumes that keys will be atomic (ie that if eg F1 is pressed
the entire sequence is present in the buffer). This is usually but not always
true, a change in the tree format so it can differentiate potential (partial)
key sequences will happens soon and will allow this to be fixed.
Convert the window pane (pty master side) fd over to use a bufferevent.
The evbuffer API is very similar to the existing tmux buffer API so this was
remarkably painless. Not many possible ways to do it, I suppose.
Initial changes to move tmux to libevent.
This moves the client-side loops are pretty much fully over to event-based only
(tmux.c and client.c) but server-side (server.c and friends) treats libevent as
a sort of clever poll, waking up after every event to run various things.
Moving the server stuff over to bufferevents and timers and so on will come
later.
partial matches to be done (they wait for further data or a timer to expire,
like a naked escape).
Mouse and xterm-style keys still expect to be atomic.
complicated because of escape input, but in that case instead of processing a
key immediately, schedule a timer and reprocess the bufer when it expires.
This currently assumes that keys will be atomic (ie that if eg F1 is pressed
the entire sequence is present in the buffer). This is usually but not always
true, a change in the tree format so it can differentiate potential (partial)
key sequences will happens soon and will allow this to be fixed.
Add an activity time for clients, like for sessions, and change session and
client lookup to pick the most recently used rather than the most recently
created - this is much more useful when used interactively and (because the
activity time is set at creation) should have no effect on source-file.
Based on a problem reported by Jan Johansson.
If it isn't available explicitly, work out the current client in a similar way
to the current session - build a list of the possibilities then pick the
newest.
Change session and client activity and creation time members to have more
meaningful names.
Also, remove the code to try and update the session activity time for the
command client when a command message is received as is pointless because it
des not have a session.
Double the escape timer (the time after a \033 is received before tmux gives up
waiting to see if it is part of a key sequence and passes it through) to 500
ms, the previous setting was too fast. Suggested by naddy.
This moves the client-side loops are pretty much fully over to event-based only
(tmux.c and client.c) but server-side (server.c and friends) treats libevent as
a sort of clever poll, waking up after every event to run various things.
Moving the server stuff over to bufferevents and timers and so on will come
later.
client lookup to pick the most recently used rather than the most recently
created - this is much more useful when used interactively and (because the
activity time is set at creation) should have no effect on source-file.
Based on a problem reported by Jan Johansson.
meaningful names.
Also, remove the code to try and update the session activity time for the
command client when a command message is received as is pointless because it
des not have a session.
Add a flag for jobs that shouldn't be freed after they've died and use it for
status jobs, then only kill those jobs when status-left, status-right or
set-titles-string is changed.
Fixes problems with changing options from inside #().
If any client currently displaying a window pane has more than 1 KB of output
buffered, don't accept any further data from the process running in the pane.
This makes tmux much more responsive when flooded with output, although other
buffers can still have an impact when running remotely.
Prompted by a query from Ranganathan Sankaralingam.
status jobs, then only kill those jobs when status-left, status-right or
set-titles-string is changed.
Fixes problems with changing options from inside #().
Rewrite xterm-keys code (both input and output) so that works (doesn't always
output the same modifiers, accepts all the possible input keys) and is more
understandable.
Support the (mostly new) function key+modifier caps (kIC-kIC7). Most of these
will be caught (soon) by the xterm keys code in xterm itself but some other
descriptions such as rxvt define them as well.
buffered, don't accept any further data from the process running in the pane.
This makes tmux much more responsive when flooded with output, although other
buffers can still have an impact when running remotely.
Prompted by a query from Ranganathan Sankaralingam.
Remove the -d flag to tmux and just use op/AX to detect default colours.
Irritatingly, although op can be used to tell if a terminal supports default
colours, it can't be used to set them because in some terminfo descriptions it
resets attributes as a side-effect (acts as sgr0) and in others it doesn't, so
it is not possible to determine reliably what the terminal state will be
afterwards. So if AX is missing and op is present, tmux just sends sgr0.
Anyone using -d for a terminal who finds they actually needed it can replace it
using terminal-overrides, but please let me know as it is probably an omission
from terminfo.
Irritatingly, although op can be used to tell if a terminal supports default
colours, it can't be used to set them because in some terminfo descriptions it
resets attributes as a side-effect (acts as sgr0) and in others it doesn't, so
it is not possible to determine reliably what the terminal state will be
afterwards. So if AX is missing and op is present, tmux just sends sgr0.
Anyone using -d for a terminal who finds they actually needed it can replace it
using terminal-overrides, but please let me know as it is probably an omission
from terminfo.
Split the server code handling clients, jobs and windows off into separate
files from server.c (merging server-msg.c into the client file) and rather than
iterating over each set after poll(), allow a callback to be specified when the
fd is added and just walk once over the returned pollfds calling each callback
where needed.
More to come, getting this in so it is tested.
Try to reduce the UTF-8 mess.
Get rid of passing around u_char[4]s and define a struct utf8_data which has
character data, size (sequence length) and width. Move UTF-8 character
collection into two functions utf8_open/utf8_append in utf8.c which fill in
this struct and use these functions from input.c and the various functions in
screen-write.c.
Space for rather more data than is necessary for one UTF-8 sequence is in the
utf8_data struct because screen_write_copy is still nasty and needs to reinject
the character (after combining) into screen_write_cell.
UTF-8 combined character fixes.
Thai can have treble combinations (1 x width=1 then 2 x width=0) so bump the
UTF-8 cell data size to 9 and alter the code to allow this.
Also break off the combining code into a separate function, handle any further
combining beyond the buffer size by replacing the character with _s, and when
redrawing the UTF-8 character don't assume the first part has just been
printed, redraw the entire line.
Instead of having a complicated check to see if the cursor is in the last
position to avoid an explicit wrap, actually move it there.
Some UTF-8 fixes to come.
files from server.c (merging server-msg.c into the client file) and rather than
iterating over each set after poll(), allow a callback to be specified when the
fd is added and just walk once over the returned pollfds calling each callback
where needed.
More to come, getting this in so it is tested.
Get rid of passing around u_char[4]s and define a struct utf8_data which has
character data, size (sequence length) and width. Move UTF-8 character
collection into two functions utf8_open/utf8_append in utf8.c which fill in
this struct and use these functions from input.c and the various functions in
screen-write.c.
Space for rather more data than is necessary for one UTF-8 sequence is in the
utf8_data struct because screen_write_copy is still nasty and needs to reinject
the character (after combining) into screen_write_cell.
Thai can have treble combinations (1 x width=1 then 2 x width=0) so bump the
UTF-8 cell data size to 9 and alter the code to allow this.
Also break off the combining code into a separate function, handle any further
combining beyond the buffer size by replacing the character with _s, and when
redrawing the UTF-8 character don't assume the first part has just been
printed, redraw the entire line.
Add mode keys to move the cursor to the top, middle and bottom of the screen.
H/M/L in vi mode and M-R/M-r in emacs (bottom of screen not bound in emacs).
When drawing lines that have wrapped naturally, don't force a newline but
permit them to wrap naturally again. This allows terminals that use this to
guess where lines start and end for eg mouse selecting (like xterm) to work
correctly.
This was another long-standing issue raised by several people over the last
while.
Thanks to martynas@ for much testing. This was not trivial to get right so
bringing it in for wider testing and adn to fix any further glitches in-tree.
When backspace is received at the beginning of a line and the previous line was
wrapped, move the cursor back up to the end of the previous line.
Another one of the forgotten persons requested this quite a while ago (I need
to start noting names on todo items...) when it was quite hard to
implement. Now it is easy and I don't see it can do any harm, so hey presto...
Use relative cursor movement instead of absolute when possible and when
supported by the terminal to reduce the size of the output data (generally
about 10-20%).
permit them to wrap naturally again. This allows terminals that use this to
guess where lines start and end for eg mouse selecting (like xterm) to work
correctly.
This was another long-standing issue raised by several people over the last
while.
Thanks to martynas@ for much testing. This was not trivial to get right so
bringing it in for wider testing and adn to fix any further glitches in-tree.
wrapped, move the cursor back up to the end of the previous line.
Another one of the forgotten persons requested this quite a while ago (I need
to start noting names on todo items...) when it was quite hard to
implement. Now it is easy and I don't see it can do any harm, so hey presto...
Add a pipe-pane command to allow a pane to be piped to a shell command, for
example:
pipe-pane 'cat >~/out'
No arguments stops outputing and closes the pipe; the -o flag toggles a pipe
and on and off (useful for key bindings).
Suggested by espie@.
Clean up by introducing a wrapper struct for mouse clicks rather than passing
three u_chars around.
As a side-effect this fixes incorrectly rejecting high cursor positions
(because it was comparing them as signed char), reported by Tom Doherty.
Rather than running status-left, status-right and window title #() with popen
immediately every redraw, queue them up and run them in the background,
starting each once every status-interval. The actual status line uses the
output from the last run.
This brings several advantages:
- tmux itself may be called from inside #() without causing the server to hang;
- likewise, sleep or similar doesn't cause the server to block;
- commands aren't run excessively often when redrawing;
- commands shared by status-left and status-right, or used multiple times, will
only be run once.
run-shell and if-shell still use system()/popen() but will be changed over to
use this too later.
New option, mouse-select-pane. If on, the mouse may be used to select the
current pane.
Suggested by sthen@ and also by someone else ages ago who I have forgotten.
Add "grouped sessions" which have independent name, options, current window and
so on but where the linked windows are synchronized (ie creating, killing
windows and so on are mirrored between the sessions). A grouped session may be
created by passing -t to new-session.
Had this around for a while, tested by a couple of people.
Support for individual session idle time locking. May be enabled by turning off
the lock-server option (it is on by default). When this is off, each session
locks when it has been idle for the lock-after-time setting. When on, the
entire server locks when ALL sessions have been idle for their individual
lock-after-time settings.
This replaces one global-only option (lock-after-time) with another
(lock-server), but the default behaviour is usually preferable so there don't
seem to be many alternatives.
Diff/idea largely from Thomas Adam, tweaked by me.
example:
pipe-pane 'cat >~/out'
No arguments stops outputing and closes the pipe; the -o flag toggles a pipe
and on and off (useful for key bindings).
Suggested by espie@.
three u_chars around.
As a side-effect this fixes incorrectly rejecting high cursor positions
(because it was comparing them as signed char), reported by Tom Doherty.
immediately every redraw, queue them up and run them in the background,
starting each once every status-interval. The actual status line uses the
output from the last run.
This brings several advantages:
- tmux itself may be called from inside #() without causing the server to hang;
- likewise, sleep or similar doesn't cause the server to block;
- commands aren't run excessively often when redrawing;
- commands shared by status-left and status-right, or used multiple times, will
only be run once.
run-shell and if-shell still use system()/popen() but will be changed over to
use this too later.
so on but where the linked windows are synchronized (ie creating, killing
windows and so on are mirrored between the sessions). A grouped session may be
created by passing -t to new-session.
Had this around for a while, tested by a couple of people.
the lock-server option (it is on by default). When this is off, each session
locks when it has been idle for the lock-after-time setting. When on, the
entire server locks when ALL sessions have been idle for their individual
lock-after-time settings.
This replaces one global-only option (lock-after-time) with another
(lock-server), but the default behaviour is usually preferable so there don't
seem to be many alternatives.
Diff/idea largely from Thomas Adam, tweaked by me.
Make C-Up and C-Down in copy mode scroll the screen up and down one line
without moving the cursor, like Up and Down in scroll mode (which will shortly
disappear).
If no target client is specified to commands which accept one, try to guess the
current client, in a similar manner to how sessions already work: if the
current session can be established and has only one client, use that; otherwise
use the most recently created client.
current client, in a similar manner to how sessions already work: if the
current session can be established and has only one client, use that; otherwise
use the most recently created client.
Support -c like sh(1) to execute a command, useful when tmux is a login
shell. Suggested by halex@.
This includes another protocol version increase (the last for now) so again
restart the tmux server before upgrading.
Remove the internal tmux locking and instead detach each client and run the
command specified by a new option "lock-command" (by default "lock -np") in
each client.
This means each terminal has to be unlocked individually but simplifies the
code and allows the system password to be used to unlock.
Note that the set-password command is gone, so it will need to be removed from
configuration files, and the -U command line flag has been removed.
This is the third protocol version change so again it is best to stop the tmux
server before upgrading.
Trim some code by moving the ioctl(TIOCGWINSZ) after SIGWINCH from the client
into the server.
This is another (the second of four) protocol version changes coming this
morning, so again the server should be killed before upgrading.
Don't attempt to open() the tty path, rely on the client sending its stdin fd
with imsg and fatal if it doesn't, then set the FD_CLOEXEC flag in tty_init
instead of tty_open to prevent them leaking into child processes if any are
created between the two calls.
This bumps the protocol version, so the tmux server should be killed before
upgrading.
command specified by a new option "lock-command" (by default "lock -np") in
each client.
This means each terminal has to be unlocked individually but simplifies the
code and allows the system password to be used to unlock.
Note that the set-password command is gone, so it will need to be removed from
configuration files, and the -U command line flag has been removed.
This is the third protocol version change so again it is best to stop the tmux
server before upgrading.
into the server.
This is another (the second of four) protocol version changes coming this
morning, so again the server should be killed before upgrading.
with imsg and fatal if it doesn't, then set the FD_CLOEXEC flag in tty_init
instead of tty_open to prevent them leaking into child processes if any are
created between the two calls.
This bumps the protocol version, so the tmux server should be killed before
upgrading.
Permit multiple prefix keys to be defined, separated by commas, for example:
set -g prefix ^a,^b
Any key in the list acts as the prefix. The send-prefix command always sends
the first key in the list.
==
Rather than constructing an entire termios struct from ttydefaults.h, just let
forkpty do it and then alter the bits that should be changed after fork. A
little neater and more portable.
==
This should fix problems caused by glibc's broken ttydefaults.h file.
use the error and exit on MSG_EXIT (it was being handled in the default
case). Undo the last change, move the errstr check into the MSG_EXIT case, and
add a comment.
use the error and exit on MSG_EXIT (it was being handled in the default
case). Undo the last change, move the errstr check into the MSG_EXIT case, and
add a comment.
When using tmux as a login shell, there is currently no way to specify a shell
to be used as a login shell inside tmux, so add a default-shell session option.
This sets the shell invoked as a login shell when the default-command option is
empty.
The default option value is whichever of $SHELL, getpwuid(getuid())'s pw_shell
or /bin/sh is valid first.
Based on a diff from martynas@, changed by me to be a session option rather
than a window option.
to be used as a login shell inside tmux, so add a default-shell session option.
This sets the shell invoked as a login shell when the default-command option is
empty.
The default option value is whichever of $SHELL, getpwuid(getuid())'s pw_shell
or /bin/sh is valid first.
Based on a diff from martynas@, changed by me to be a session option rather
than a window option.
Add a new display-panes command, with two options (display-panes-colour and
display-panes-time), which displays a visual indication of the number of each
pane.
Add a choose-client command and extend choose-{session,window} to accept a
template. After a choice is made, %% (or %1) in the template is replaced by the
name of the session, window or client suitable for -t and the result executed
as a command. So, for example, "choose-window "killw -t '%%'"" will kill the
selected window.
The defaults if no template is given are (as now) select-window for
choose-window, switch-client for choose-session, and detach-client for
choose-client (now bound to D).
template. After a choice is made, %% (or %1) in the template is replaced by the
name of the session, window or client suitable for -t and the result executed
as a command. So, for example, "choose-window "killw -t '%%'"" will kill the
selected window.
The defaults if no template is given are (as now) select-window for
choose-window, switch-client for choose-session, and detach-client for
choose-client (now bound to D).
When using source-file, run the commands in the context of the source-file
command rather than with no context. This makes things like attach work from a
file.
Extend command-prompt with a -p option which is a comma-separated list of one
or more prompts to present in order.
The responses to the prompt are replaced in the template string: %% are
replaced in order, so the first prompt replaces the first %%, the second
replaces the second, and so on. In addition, %1 up to %9 are replaced with the
responses to the first the ninth prompts
The default template is "%1" so the response to the first prompt is processed
as a command.
Note that this changes the behaviour for %% so if there is only one prompt,
only the first %% will be replaced. Templates such as "neww -n '%%' 'ssh %%'"
should be changed to "neww -n '%1' 'ssh %1'".
From Tiago Cunha.
Instead of just checking for an empty buffer, which may not be the case if
there is unconsumed data, save the previous size and use it instead. This means
that activity monitoring should work in this (unlikely) event.
Also remove a debugging statement that no longer seems necessary.
