When Q2 runs in windowed mode, SDL can never switch to the wrong the
resolution. The resolution isn't switched, of course. The work around
prevented Q2 from creating Windows larger than the resolution of the
primary display. For example a primary display of 1280x1024 prevented
a window size of 1680x1050 on the much bigger secondary display.
Normally stdout and stderr are buffered. In case of a crash the last
lines aren't written to the stdout.txt, making post mortem debugging
difficult. Forcing both FDs to unbuffered mode ensures that everything
gets written. The performance impact is negliable.
Even tough it is a multi user os, due to BeOs heritage and being backwards
compatible with, the desktop runs as root, thus its dog get eaten after all...
Combs all Raw search paths to find game dirs containing PAK/PK2/PK3
files. If multiple uniquely-named directories exist, then show a "mods"
option on the "Game" menu and allow selection of desired mod on new
eponymous submenu. Includes fix for memory leak of mapnames (read from
"maps.lst") when changing games.
When the "game" directory is changed, clear the current list of maps in
the "start network server" menu so that it will be re-initialized the
next time the menu is accessed.
so it waits for about the time of one frame at 60fps, but independently
of the actual framerate.
Without this fix, wait is broken unless vsync is on, because
CBuf_Execute() is called about 1000 times per second by Qcommon_Frame(),
even if no render- or packet-frame is executed.
(vsync "fixes" this because then we have a real wait at the end of each
renderframe)
We busy-looped for 5000 microseconds, i.e. 5 milliseconds, which reduces
the framerate to < 200fps
I guess the value was copypasted from Sys_Nanosleep() below, but
that was nanoseconds of course..
Anyway, busy-looping for 5 *micro*seconds instead fixes it.
Angle kicks are used for weapon recoil, damage kicks and the like.
Setting the `cl_anglekicks` to `0` ignored them, they're read from
the network but not displayed. The cvar is cheat protected, it's
reset to `1` on server connect.
Rquested by Martin via mail.
In coop it's often hard to get on the same elevator together, because
they're immediately triggered once the first player steps on it.
This cvar sets a delay (1 second by default) for the elevator to wait
before moving, so other players have some time to get on it.
If you like elevators/platforms that suck, just set it to `0` :-P
Currently only used in func_plat, if it turns out that other entities
are used for automatically triggered platforms, we'll have to adapt
those as well (I guess wait_and_change() is generally useful for that).
We're not bumping the savegame version because they should only break in
an uncommon corner case: *Coop* savegames created with clients including
this change will not work on older clients - SP savegames are not
affected and old savegames on new clients also still work.
In coop a weapon can be picked up only once. That's annoying, because in
coop ammunition is sparse and not getting the ammunition that comes with
a weapons make things worse. When `coop_pickup_weapons` is set to `1` a
weapon may be picked up if:
1) The player doesn't have the weapon in their inventory.
2) No other player has already picked it up.
This is one these constructs which makes you wonder how it could ever
work. When querying a cvar by calling Cvar_Get(), the default value
(given in `var_value`) is copied into `cvar_t->default_string`. If a
NULL pointer is given in `var_value`, the NULL pointer is passed to
CopyString() and dereferenced. The game crashes. There's already a NULL
pointer check in the 'cvar wasn't found' branch, but none in the 'cvar
was found' branch... Moving the check to the beginning of the function
isn't an option, because at least lithium2 doesn't implement a NULL
pointer check either. We would just move the crash from the server into
the game.dll. Therefore copy an empty string into
cvar_t->default_string` when a NULL pointer was passed in `cvar_value`
and the cvar was found. Pass the empty string trough `CopyString()` to
get an Z_MAlloc() allocation for it, otherwise we would call `Z_Free()`
on an unallocated object further down below.
Reported by Chris Stewart.
ogg_enabled doesn't prevent music from playing, it just toggles if the
ogg backend should be enabled or not. If the user does something like
`ogg_enable=0; snd_restart` everything is okay. If they just set
`ogg_enable=0` strange things happen because the backend stays
initialized and will play tracks as requested. Work around the by
shutting the backend down if `ogg_enable == 0 && ogg_started == true`.
Closes#583.
This allows really bug configuration files up to 32k. It would be better
to switch the global buffer to something allocated at runtime, but thats
non-trivial... This change should be save, since the buffers are global
(allocated in the BSS) and not included in savegames nor send over the
network.
Closes#582.
It's an often reported, that the q2ded dedicated server consumes huges
amounts of CPU time. That's because users don't know that `busywait`
must be set to `0`. Since there's no point in using busywaits in the
dedicated server (the network jitter is always bigger than the
jitter caused by nanosleep() and equivalents), just force q2ded to
use nanosleep().
Some projectiles like grenades or rockets are classified as enemies.
Their explosion spawn a sound entity, monsters should move to that. But
the projectile is destroyed when exploding, it's entity struct is set to
NULL. Therefor the self->enemy pointer is also NULL. The self->enemy
check was removed in bc5f5698. Work around this by pretending that the
enemy is already there.
This was reported by @Soldy, closesyquake2/xatrix#56.
This is a special case, the bug was already present in Vanilla Q2: If a
server is offering assets for download but is missing some files the
USP download code runs in an endless loop. CL_ParseDownload() detects
that something is wrong and calls CL_RequestNextDownload() which tries
to download the same file again... Work around this by skippig over that
file.
This closes#552.
The new ordering was introduced in 16ee007, fixing some problems with
the wrong sound getting played when an entity triggers several sound at
the same timestamp. This broke the behavior of the mods, in #558 dday
was mentioned, muzzle flashe sound prevent the firing sound from getting
played.
Since we don't control the source of all mods, add a simple band aid
fix: Use the new ordering for baseq2, xatrix and rouge. Use the old
ordering for everything else.
An alternative approach is being discussed in #564.
I don't remember why we restricted it to client startup. The original
code executed it everytime when `game` changed... Revert to that
behavior. Look here if some problems come up. ;)
Closes#544.
Adaptive vsync is a often requested feauture and easy to implement. Set
`r_vsync` to `2` to enable. This is untested because my system doesn't
support it.
There already was one (that I only recently fixed) for semicolon, but
the same problem can happen with quotes or $ (which is used in macros)
(single-quote ' is probably not affected, added it just to be sure)
regardless of keyboard layout, with a special exemption for layouts
where that key generates a quote character (like the Brazilian one)
because you may wanna type a quote into the console.
(It's SDL_SCANCODE_GRAVE, that key between Esc, 1 and Tab)
The old hack of matching for ^, ~ and ` in Char_Event() didn't work very
well for layouts we didn't anticipate, which is especially relevant with
the recent Scancode fallback, which for example allows binding the ^ key
on Belgian keyboards (which is on SDL_SCANCODE_LEFTBRACKET, far away
from the "console key"), but in that case would *also* open the console.
This is mostly straight-forward, except for a small hack to prevent the
key from generating text input (on German layouts you otherwise get
"^" in the console when closing+opening it), which requires the
"console key" to be pressed without any modifiers like Shift or AltGr.
Yes, it's ugly, but it works and all the uglyness is contained in
IN_Update() and on the other hand Char_Event() becomes less ugly :)
If we can't map a SDL_KEYDOWN/KEYUP events SDL_Keycode to a known
Quake2 K_* keycode, we try to map the SDL_Scancode to one of the new
K_SC_* YQ2 scancodes instead.
The scancode name corresponds to the key at that position on US-QWERTY
keyboards *not* the one in the local layout, for example the German 'Ö'
key is K_SC_SEMICOLON.
This way (hopefully!) all keys on common keyboards can be bound,
regardless of their layout. The key name won't be immediately obvious
to the user, but it's only a fallback and better than nothing.
fixes#543
this way it's easier to tell if a key constant is not handled.
Also, there was a half-finished workaround to allow binding a ';' key
(which apparently in configs would otherwise be interpreted as
command separator), now that should actually work (=> special case
it in Key_KeynumToString())
This is an enhancement to the previous `yield` work:
* Don't enforce `-march=armv8-a` for aarch64 builds, because it is the
initial ARMv8 revision and compilers will either use that or something
newer.
* Refine preprocessor guards around `asm("yield");` so the code isn't
compiled in if unsupported by the current `-march='.
Submitted by @smcv in #535.
On MacOS texture is cleaned up after render and code have to copy a whole
screen to texture, other platforms save previous texture content and can
be copied only changed parts.
YQ2 has a much more precise Sys_Milliseconds() than Vanilla Quake II and
it always start at 0, not some other semirandom value. If the client is
started by `./quake2 +connect example.com" or all user just walk their
way to the menu there's a very high propability that two ore more
clients end up with the same qport... We can't use rand(), because we're
always starting with the same seed, so all clients generate more or less
the same random numbers and we end up in the same situation.
So just call time(). It's portable and more or less in line what the
original code did for Windows. It may be necessary to implement some
kind of fallback logic just in case that still two clients end up with
the same qport, but that's a task for another day.
Closes#537.
This was introduced in 220f0a9 as fix. The submitter, @DanielGibson and
myself missunderstood the code:
* If net_interface is NULL (which in the current code can never happen),
an empty string (the user sets the `ip` cvar to an empty string) or
"localhost" (the default) we want to set `Host` to the unspecified
address. getaddrinfo() will return in6addr_any fot it and we'll bind
to any available address.
* "0.0.0.0" isn't the IPv4 any address, it's the unspecified address.
Thats correct and the code was working fine for IPv4. But at least
the submitter and me confused it with the any address (which is
0.0.0.0/0). So setting `Host` in the IPv6 to `::/128` (the lowest IPv6
address) or `::/0` (any IPv6) is wrong, it must be `::` (unspecified
IPv6 address)!
Have a look at RFC 3493 for the details.
I'm doing the change only for the Unix code path, not for Windows. For
some reason everything besides `::/128` or `::1` doesn't really work on
Windows and I don't know why. Even more scary is that changes to the
IPv6 case also break IPv4 sockets. Since the whole network.c for Windows
is confuse and rather hard to understand (there's still IPX support in
it) I'm leaving things as they are.
C11 _Noreturn is only accepted on function declarations, not on function
pointers, so we can't use it on callbacks like game_import_t.error and
refimport_t.Sys_Error. Use a separate macro for those.
The problematic situation doesn't currently happen because the Makefile
hard-codes -std=gnu99, which disables C11 features; but removing
-std=gnu99 (resulting in the compiler's default, currently gnu11) causes
compilation failures with at least gcc 9.x.
Signed-off-by: Simon McVittie <smcv@debian.org>
Until now CFGDIR was hardcoded to YamagiQ2 on Windows and .yq2 on
everything else. Sometimes it's desireable to have a separate dir
for some tasks, for example whentesting things that introduce new
cvars. Add -cfgdir to override CFGDIR.
This allows for longer arguments passed to cvars, gl_nolerp_list is a
good example for a case were a token length of 128 is too short. Keep
the mapname[] buffer in the server struct at 128 bytes to preserve
savegame compatibility.
Closes#526.
In the vanilla code show_hostile was a qboolean what's clearly wrong.
For wome reasons I don't remember I changed it to an integer and added
the casts. This is problematic because show_hostile is derived from
level.time which is a float. The loss in precision broke some corner
cases like monsters becoming activated when they shouldn't.
Found, analyzed and reported by @BjossiAlfreds #525. Closes#525.
Until this commit a cinematic was aborted as soon as any key were
marked down when finishing the user command and sending it to the
server. The whole logic to detect if a key is down is broken, for
example `vid_restart` may leave keys marked down that are in fact
up. And there's the possibility to inject fake key events from
nearly everywhere. I'm not really sure but I suspect that even the
server may be able to inject key events.
Therefore untangle the cinematic abort code from the user command
processing, it should depend only on real key strokes:
1. Introduce a new global variable `abort_cinamatic` and set it to
`cls.realtime` as soon as a key down event is detected. The only
exceptions are Escape and Shift, because opening the menu and
toggeling the console should never abort a cinematic.
2. When starting a cinematic `abort_cinamatic` is set to INT_MAX,
because it needs to be higher than the current `cls.realtime`.
3. When a cinematic is running, `cls.key_dest` is set to `key_game`
(`key_menu` and `key_console` are ignored, keys send to the menu
or the console should never abort a cinematic; `key_message`
can / should never happen while a cinematic is running) and
`abort_cinamatic` is less than `cls.realtime` the cinematic is
aborted.
`abort_cinamatic` less than `cls.realtime` is necessary because the
client needs one frame to pop up the menu or toggle the console and set
the `cls.key_dest` accordingly. `abort_cinamatic == cls.realtime - 1`
is not possible because not every frame finishes a user command.
This closes#502.
An option like this was often requested since I fixed the gun field of
view in e466554. Since the software renderer is missing the ability to
alter the persepective matrix (e.g. something like glFrustrum()) this
fix fakes the offset by manipulating the guns transformation marix.
That's not perfect, the gun distorts if `sw_gunzposition` is set to
anything but `0`. Values up to `8` are more or less okay. Defaults
to `8` which matches the GL renderer default.
Pushing all entities slightly away from non-horizontal may let items to
slide to unreachable locations, or let monsters getting stuck.
This is part of yquake2/xatrix#50
