This commit adds two new configuration options: igpu_desiredgov and
igpu_power_threshold which allow for a different CPU governor when the
Intel integrated GPU is under load. This currently only applies to
Intel integrated GPUs and not AMD APUs because it uses the Intel RAPL
infrastructure for getting power information. If on a platform that
without an Intel integrated GPU or where the kernel does not support
RAPL, the new options will be ignored and it will fall back to the old
behavior.
One of the core principals of gamemoded to date has been that, when
playing a game, we want to use the "performance" CPU governor to
increase CPU performance and prevent CPU-limiting. However, when the
integrated GPU is under load, this can be counter-productive because the
CPU and GPU share a thermal and power budget. By throwing the CPU
governor to "performance" game mode currently makes the CPU frequency
management far too aggressive and it burns more power than needed. With
a discrete GPU, this is fine because the worst that happens is a bit
more fan noise. With an integrated GPU, however, the additional power
being burned by the CPU is power not available to the GPU and this can
cause the GPU to clock down and lead to significantly worse performance.
By using the "powersave" governor instead of the "performance" governor
while the integrated GPU is under load, we can save power on the CPU
side which lets the GPU clock up higher. On my Razer Blade Stealth 13
with an i7-1065G7, this improves the performance of "Shadow of the Tomb
Raider" by around 25-30% according to its internal benchmark mode.
SCHED_ISO is not supported by upstream kernels, so don't try to use this by
default since for most users it will result in an error log that trying to
set the scheduling policy fails.
Without extra system configuration, we will also not have permission to renice
processes by default, so out of the box doing this will fail as well.
Users that wish to use these features can enable them from the config once
they have configured their system appropriately.
This also fixes the instances in testing where we don't have the nv overclock in use, but we do have the mode set
Solves issues explaining the what the perf_level actually meant, and future proofs for any PR that wants to set individual perf levels
This covers the MVP for now, and simply allows pinning the power level to "high"
Full overclocking set up is somewhat more complicated, and it'll be better to implement that at the same time as the same for Nvidia, where we're currently only really setting the top end power level
This allows direct control over who can make requests on behalf of other processes
require_supervisor can also be used to allow a supervisor to take direct control of gamemode on the system (perhaps a GUI, or game launcher)
This commit adds configuration support for the renice value and amends
documentation and examples. This commit by itself does nothing, the
following commit is needed to actually apply the new settings.
Signed-off-by: Kai Krakow <kai@kaishome.de>
This adds support for a new configuration option "softrealtime" to be
read from the general section. This commit alone does nothing, the
following commit adds actually making use of the value.
Signed-off-by: Kai Krakow <kai@kaishome.de>
As seen in Issue #45 (thanks OlliC)
This can be placed in ~/.local/share/applications/ to allow easier (re)launching of gamemode, in particular when the DISPLAY param is needed
* Minor C cleanup
- some symbols can be made static:
1. set_gov_state
2. everything in gamemode_client.h
- daemonize() can also take a const char*, since the name is only
passed to printf() or syslog()
- prevent shadowing of variables
- use explicit (void) as parameter-list more consistently
- use some more const.
Move cast to more appropriate place and document that execv() behaves
as if args where of type const *char and we trust on that.
- example: Just use main(void), which is also an acceptable ISO-C decl
- example: Use stderr for errors
* Fix -Wold-style-declaration issue
as it's already required by meson.
Also use the default unit file path (/usr/lib/systemd/user/) for
packaged versions. /etc/systemd/user is intended for manually installed
services.
arch-meson is just a Meson wrapper with correct arch packaging prefixes set.
Updating the README.md to bring it ready for wider release:
* Removed some points that were more implementation details
* Added Contributions section
* Simplified the usage section
* Expanded initial explanation
* Fixed steam launch command to allow the overlay to still work
Others:
* Mark as 1.0
* Update the license file for 2018
Checks for a gamemode.ini in /usr/share/gamemode/ (or in the cwd for debugging)
Currently allows for blacklisting and whitelisting clients based on rudimentary needle-haystack executable name checks
See the example/gamemode.ini file for expected syntax
Using the BSD licensed inih library (with additional meson.build file)
This exposed a bunch of issues that needed dealing with to ensure the
code is clean and sane. Notably the dlopen/dlsym routine has been altered
to closer match the LSI approach of safe symbol binding, by not attempting
to directly cast the result of a dlsym operation. Instead, if we succeed
in getting the dlsym() pointer, we memcpy this to the target and ensure
we have the correct constraints.
Note that in sanitizing the log helpers, I opted to remove the varargs
ability from FATAL_ERRNO given this is used exactly like perror() and
there are no examples currently using varargs with this in the tree.
This allowed me to keep the log helpers as macros and not have to implement
wrapper functions.
Signed-off-by: Ikey Doherty <ikey@solus-project.com>
Primarily we convert the service into a thread safe one that isn't reliant
on signaling for control flow, eliminating data race conditions. We also
enable interleaving by separating game mode pivoting from explicit client
registration.
The static pid list is now converted into a dynamic list that is OOM safe
to store all registered clients (with a reasonable upper limit of 256 clients)
to better handle cases where LD_PRELOAD is used for a large process group.
Additionally we begin storing some metadata on the connected clients such
as their executable path, which will enable us to perform some basic
whitelisting in future.
The cpugovctl binary is now moved into the libexecdir as an explicit helper
of the D-BUS service, using the shared library to merge some code back into
the daemon. This saves having to execute a process to query the state of the
governors, as we don't need a privileged client to do this.
In order to sanely set the governors, we require that the binary is running
as euid 0, and execute this using `pkexec`. A PolKit policy definition is
provided which allows active/logged in users to execute this helper through
a path whitelist. As such we can convert the daemon into user-mode only, with
the privileged helper being dispatched exclusively via polkit. This removes
the need for a setuid helper or having a system mode daemon.
Lastly we clean up the codebase a bit to be consistent with modern C code
conventions, using pragmas where available. The library component still uses
the older ifdef approach to support older compilers, but the daemon portion
uses the directive to simplify intent and speed up compilation. Additionally
we move all comments to C style comments for consistency, instead of mixing
in C++ style single line comments, in order to establish a formal coding
style.
The net result is a more robust service which can be D-BUS activated when
clients need it, that can perform scaling automatically without harassing
the user with authentication popups.
Signed-off-by: Ikey Doherty <ikey@solus-project.com>
This also explicitly constructs reusable library components to make it
easier to extend the project over time. Notably we make less assumptions
about the host system and use pkgconfig where appropriate to give us
system details, such as the systemd system unit directory for packaging.
This can be overriden with the meson option.
Signed-off-by: Ikey Doherty <ikey@solus-project.com>
