scx_flash: drop user lock priority boosting #1444
Conversation
Drop the logic that tries to track user-space futexes and extends task time slices when tasks is holding a user-space lock. This improves scheduler portability and maintainability across kernel versions. For the scheduler's primary use case (audio/multimedia), the user lock priority boosting provided minimal benefits and sometimes negatively impacted performance predictability by over-prioritizing lock-heavy tasks. Moreover, conditionally load the remaining kprobe (vfs_fsync_range) to improve scheduler portability and allow to run the scheduler also in kernels where this function is not available. Signed-off-by: Andrea Righi <arighi@nvidia.com>
arighi
force-pushed
the
flash-drop-user-lock-boost
branch
from
0bdd32e to
ad3d674
Compare
likewhatevs
left a comment
likewhatevs
left a comment
There was a problem hiding this comment.
Nice, also I'm gonna move a wrapper around copies of that that if-else-unsafe block that are in layered also to a helper in compat which you can use some time later today/tomorrow btw.
|
@likewhatevs I was also wondering why in layered you're not using compat::ksym_exists() and you check /proc/kallsyms directly. Is it because nvidia_open() is in an external module? In that case we should probably fix compat. |
Just confirmed compat doesn't work for nvidia symbols (because they're a module I think, but not certain of that). From a system w/ nvidia drivers loaded (where the kallsyms approach works): |
Drop the logic that tries to track user-space futexes and extends task time slices when tasks is holding a user-space lock.
This improves scheduler portability and maintainability across kernel versions. For the scheduler's primary use case (audio/multimedia), the user lock priority boosting provided minimal benefits and sometimes negatively impacted performance predictability by over-prioritizing lock-heavy tasks.
Moreover, conditionally load the remaining kprobe (vfs_fsync_range) to improve scheduler portability and allow to run the scheduler also in kernels where this function is not available.