We posted on X a few days back, calling Linux 7.3 one of the most performance-focused releases in years. And now, going through what has actually landed in the merge window since then, that claim does seem to hold up.The merge window for this release opened right after Linux 7.2 arrived, staying open until the end of this month barring any last-minute happenings that move the closure.So, here's what is confirmed to go into the 7.3 release so far.Storage gets a serious speed boostBtrfs picks up several changes at once. Direct I/O now runs through an IOmap bounce buffer instead of falling back to buffered I/O, taking a workflow that topped out around half of the theoretical maximum throughput up to close to 95 percent of it.Two separate fixes in the logging and fsync paths add roughly 5x gains each, one from removing an unnecessary delay when multiple logging tasks run at once, the other from skipping unneeded hole detection during full fsync on files with lots of extents but no holes.The default free-space cache also moves from v1 to v2, the safer and faster option that had not yet become the default.EXT4 gets its own round of work too. Parallel direct I/O writes are now allowed where the filesystem used to be too restrictive, and the multi-block allocator skips unnecessary work through fallocate().Underneath both filesystems sits a reworked IOmap layer built around a single iterator callback the compiler can inline instead of dispatching indirectly.You don't need to do anything for any of this. It just works once you're on 7.3, and the IOmap change pays off the most on fast NVMe, pushing EXT4 from 1.92 million IOPS to 2.19 million in testing.Important scheduler workPeter Zijlstra's flatten-the-pick patches rework cgroup scheduling around a single runqueue instead of the old, hitch-prone design. Peter tested it on an aging Sandy Bridge box paired with an AMD Radeon RX 580, and both frame rate and frame-time consistency improved.If you're gaming on hardware that's seen better days, this is the one improvement that should make it feel less choppy.Intel hybrid CPUs get fixed too. Cluster-aware scheduling has not handled the P-core and E-core mix well, and Ricardo Neri's work closes that gap on Alder Lake, Lunar Lake, and Panther Lake.Practically, that means less stutter when your laptop is juggling something heavy like a source build, a game, or a video export, while background tasks compete for the small cores.Memory management fixesZTE engineers found a bug in KSM's reverse-mapping, the step where the kernel checks which memory areas point to a merged page before touching it. Under heavy load, one internal lock could stay stuck for hundreds of milliseconds, freezing other things trying to use it.The fix drops the worst-case from around 700ms to under 2ms. Simply speaking, it means fewer random freezes on servers and containers under memory pressure.Xiaomi engineer Wenchao Hao fixed a similar slowdown in zsmalloc, which zRAM and Zswap use to store compressed memory.A Raspberry Pi 4B saw gains up to 1.83x in testing, a 20-core Intel system up to 1.4x. In plain terms, you get snappier performance on budget boards like the Pi and a smaller but real gain on beefier hardware.Lower latency for real-time workloadsEvery time one CPU core uses the kernel's SMP code to ask the others to run something right away, like flushing memory caches, the kernel used to freeze everything else on that CPU until all the other cores responded. ByteDance engineers traced the slowdown to that wait and reworked it so other tasks can keep running instead of sitting frozen. On busy systems, that frozen wait could stretch out and stall high-priority tasks.In fleet testing with DPDK, a framework built for fast networking, the worst-case delay for the slowest requests dropped from ~17ms to ~1.5ms. Most desktop users will not notice this one, but for networking and real-time software, it is an important fix.Some smaller changesRISC-V adds support for eleven newer ISA extensions.EFI runtime calls now time out instead of stalling indefinitely.RAID 5/6 gets a better benchmark-based algorithm selection at boot.New AES library APIs clean up how the kernel handles common encryption modes.Wrapping up. As you might have read recently, Linus Torvalds spent 18 kernel boots and 24 debug patches chasing a one-line Intel Xe driver bug. AI helped him along the way, and the fix is already merged, of course.The merge window for Linux 7.3 will stay open for a few more days before the release candidates start rolling out, so there's still room for more to land, and the stable release is expected sometime in October.