A new EV_ET flag allows ev_fd sources to operate in edge-triggered instead
of level-triggered mode. See epoll(7) for more information.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We used to disable event sources on EV_HUP to avoid draining the
event-queue for non-edge-triggered sources. However, this turns out to be
not working with special files so we disable it again.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
This new helper can be used to check whether a display has a pending
vsync'ed buffer-flip scheduled.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
The uterm_screen API is no longer used internally and we do not plan to
extend it. So we can use uterm_display as a replacement to make the API
thinner and easier to use.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Instead of using the uterm-screen indirection, we now directly access
uterm-display objects. We do not really intend to use virtual screens with
kmscon so there is no need to make this more complex.
I don't think consoles should every provide this feature. Instead, you
should use real compositors for such tasks like Wayland+wlterm. kmscon,
however, is rather an emergency tool or a safe backup than a fancy
daily-use-console.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If Shift+Tab is pressed, most keyboard layouts map this to ISO_Left_Tab
which itself should be converted to \e[Z.
Reported-by: Vladimir Kravets
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
This implements VT switching for cdev sessions. This allows to run
external applications as sessions inside of kmscon.
This should only be used for backwards compatibility. There is really no
reason to develop new applications to do this. Instead, use
system-compositors which make all the work for you!
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
This option specifies a shortcut that can be used to schedule the dummy
session on the given seat.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Two new helpers to schedule a specific session or a session given by a
hint-number. These can be used by other subsystems to schedule
their/others sessions.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
This really doesn't make sense. We no longer want the applications to be
in charge of VT switches so always force the VT-master to be in control of
the VT-switching logic.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We always considered the dummy session to be no real session. However,
under special circumstances it might be desirable to schedule the dummy
session (very important for debugging). Therefore, change the logic to
allow scheduling the dummy session.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We need to correctly put input devices asleep only if we really disable
the VT. If our callbacks fail and prevent the VT-switch, we must keep the
input devices awake. Otherwise, we will lose control over the
sleep-states.
Also correctly put the device asleep during shutdown.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Use "const" even for void pointers as gcc will complain otherwise. Source
arguments are read-only, anyway, so this doesn't break anything.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Print debug messages when doing asynchronous session switches and handle
EINPROGRESS as special error code.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we switch VT and the current session cannot be put asleep, and later it
notifies us that it went asleep, we should then switch VT and not
reschedule the sessions.
We do this by remembering whether a VT switch is pending. The new
uterm-retry infrastructure allows us to easily repeat VT switches then.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we force a seat to go asleep, we should _always_ notify the parent
about this. Otherwise, we might get unexpected results.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Two new helpers allow sessions to mark them as background and foreground
sessions. This will immediately switch them into the new state.
The internal infrastructure was already available, but we currently always
used foreground sessions.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
A new helper allows sessions to manually handly input events so the core
does no longer perform session/VT switches on their behalf.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we cannot immediately perform a VT switch away from our application, we
might want to retry this shortly after. Therefore, we can now call
uterm_vt_retry() which then performs any outstanding VT-switches.
To avoid unexpected behavior, we limit this to a 2-3 second delay.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If the user presses some keys to switch VTs and the target VT is our VT
(which is already active in this case), we shouldn't do anything. Calling
VT_ACTIVATE doesn't particularly hurt, but it doesn't get us anything
here so skip it.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we deactivate our VT and the target-VT is already active (probably
because we opened the _current_ VT during uterm-vt creation), we shouldn't
do anything but simply close the VT.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we force a pause we need to forward the error code. Otherwise, we
cannot properly react on it and print warnings.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we are asleep and in foreground, go_foreground() should return success
so we can safely unload asynchronous sessions. Otherwise, we wouldn't be
able to handle them properly while being inactive or we would need a whole
bunch of other checks.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we want to run as client under user-space fake-VT implementations like
kmscon-cdev, we need to be able to catch those signals from
userspace-daemons. Therefore, do not check whether the signals were
generated by the kernel.
This doesn't affect security in any way as non-priviledged processes
aren't allowed to send signals, anyway. Moreover, this increases
debugability of VT layers a lot as we can fake these signals now.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
The logic behind this two-layer check is bogus and doesn't work. Merge
both layers so we have a proper dependency-check again. This will make the
error messages less verbose but at least it works again.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We cannot allocate VT objects without an associated input object.
Therefore, allocate a dummy device with no input devices assigned to it
and use it for the test-vt object.
Note that this makes switching from/to this VT via command-line
impossible.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We cannot use 'ret' in error paths, otherwise, we might return 0. This
might then cause instabilities as the objects weren't initialized
correctly.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We do not use the TTY for anything except VT functions. Hence, there is no
need to modify the TTY settings before using it.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If a session is deactivated asynchronously, it can now notify the seat
when it is done. This will then cause a re-schedule of the session list.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we want to allow external programs to control sessions inside of
kmscon, we need to give them some time for session
activation/deactivation. Therefore, the whole session-management is
asynchronous now and allows the session code to return EINPROGRESS or
similar if they need more time.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we iterate over a non-empty list and pass the same pointer for start
and head, then we will never leave that loop.
It might be unclear why this is supported at all as we can simply use the
shl_dlist_for_each() for that. However, it makes code a lot easier to read
if the user can simply use shl_dlist_for_each_but_one() without checking
whether they passed the same for start and head.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Make the gltex renderer use bold fonts for bold characters. Internally, we
use a separate hashtable to cache all glyph information. As the hash-key
is the unicode ID, we have to use a separate table for bold glyphs. This
doesn't affect performance, though.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
This makes the terminal layer allocate bold fonts (and use normal fonts as
fallback) and pass it to the text layer. This allows the text layer to
actually draw bold characters.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
We now allow the user to specify the font that is used for bold
characters. If NULL is given, the normal font is used.
The bold font is not used by any renderer backend, yet, but will be hooked
up later.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
This almost certainly can be optimized, but we should be able to mask
getty@.service with kmscon@.service.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Fix the problems caused by the keyboard state getting out-of-sync with
the xkb_state while the device is asleep (instead of the current hack).
The problem is described in a comment and in the removed FIXME.
Signed-off-by: Ran Benita <ran234@gmail.com>
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
The current code ignores non-keyboard devices, and various places
implicitly assume that it does. If we ever need finer granularity, we
should review (and probably restructure) the code for that; but for now
this test seems out of place.
Signed-off-by: Ran Benita <ran234@gmail.com>
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
eloop does already block signals that we register callbacks for so we do
not have to block it manually here.
Note that the signal-handlers are registered _before_ doing the backend
initialization so we also avoid the previously fixed race-condition.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we open a VT, we assume that we are the only user on it. Therefore, we
put it into graphics mode _once_ and keep it this way. Resetting it to
text mode is useless, as there is no other user and every VT has its own
modes.
This also fixes page-flip issues during VT-switches which caused the
text-mode fbdev layer to claim one vertical-sync for their own page-flip
only for us to override it immediately after it is done.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
Instead of doing the backend initialization directly in uterm_vt code, we
now use fake_* helpers similar to real_* helpers.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
If we dispatch VT-deactivation to child-processes, we might allow them to
abort the VT-switch, so the upper layer (uterm_vt) must also allow us to
abort the VT-switch.
During shutdown, we need to force the deactivation to guarantee that a VT
is inactive during shutdown.
Signed-off-by: David Herrmann <dh.herrmann@googlemail.com>
