# Linux 7.3 released with GPU VRAM management, Btrfs/Ext4 gains, and sched_ext completion

> **Key takeaway:** Linux 7.3 brings completed sched_ext BPF-based cgroup sub-schedulers, improved VRAM contention handling for low-end GPUs, and significant Btrfs and Ext4 filesystem performance improvements.

**Published:** 2026-10-19T17:01:56+00:00 
**Source:** CNX Software 
**Category:** open-hardware 
**Canonical URL:** https://fosswire.org/news/linux-73-released-with-gpu-vram-management-btrfsext4-gains-a.html 

## Summary
Linus Torvalds released Linux 7.3 on schedule, with the merge window for 7.4 opening immediately after. The release completes the sched_ext implementation for BPF-based per-cgroup task scheduling. Filesystem improvements include Btrfs Direct I/O performance recovery and up to 5x faster fsync with no-holes, plus Ext4 direct I/O parallel write and fallocate() gains. Low-end GPUs benefit from dmem cgroup-based VRAM eviction to reduce application ping-ponging. FUSE io_uring gains zero-copy and buffer pool support. Architecture work spans Allwinner A733, Rockchip RV1106 and RK3576/RK3588 boards, Amlogic A1/T7/A9, Samsung Exynos 8855/1580 and ExynosAuto v920, and Qualcomm's new 'Shikra' Cortex-A78C SoC.

## What's in Linux 7.3 for developers and embedded
- **sched_ext completion**: cgroup sub-schedulers now fully functional, allowing per-cgroup BPF-based task schedulers with different policies per cgroup
- **GPU VRAM**: dmem cgroup enables eviction of unprotected buffers instead of immediate backoff when applications compete for limited VRAM
- **Btrfs**: Direct I/O performance restored; 1 jiffy delay removed for non-SSD logging workloads yielding 5x throughput increase; fsync with no-holes (default since btrfs-progs 5.15) 5x faster
- **Ext4**: direct I/O parallel writes and fallocate() improved; buffered-write regression fixed
- **FUSE io_uring**: kernel buffer pools reduce memory requirements; zero-copy path added
- **Arm/RISC-V/MIPS**: extensive SoC and board support additions including Rockchip RV1106, RK3576, RK3588 variants; Qualcomm 'Shikra'; Samsung Exynos 8855/1580, ExynosAuto v920; Allwinner A733

## Why it matters
The completed sched_ext enables production use of programmable per-cgroup scheduling, potentially improving latency-critical workloads. Embedded and SBC users gain substantial new hardware support across Rockchip, Qualcomm, Samsung, and Allwinner platforms.