# Linux 7.4-rc2 arrives with FAT32 up to 162 times faster

> Torvalds releases the second kernel release candidate with a preallocation patch for FAT32, memory fixes reported by Anthropic and a signal in BPF.
> By José Maldonado · Technology · 2026-11-10

Linus Torvalds released the second release candidate of Linux 7.4 on the night of Sunday, November 8, eight days after an rc1 that he described as the largest in the project's history: nearly 20,000 commits compared with the usual 15,000. The stable version is expected by the end of December and is being tipped as a candidate to become the next long-term support kernel.

The list of actual changes is more modest than its size suggests. Much of the volume consists of automatic Coccinelle conversions that Torvalds received during the merge window and held back while he was traveling, and a similar share are dumps of AMD GPU logs. Once that noise is removed, the kernel splits into three thirds.

[The split is one third per side:](https://www.linuxcompatible.org/story/linux-kernel-74rc2-released-largest-merge-window-in-history-meets-stabilization-test/)drivers (GPU, network and sound), networking and BPF, and the rest, which includes file systems, memory management and architecture changes. [At BPF, fixes arrive faster than new reports](https://www.webpronews.com/linux-7-4-rc2-torvalds-bpf-ai-bug-reports/), which is the part of the announcement that is circulating the most. That detail, which may seem minor, is the first sign that the tide of patches generated by automated bug detection is starting to get a response from the other side.

## FAT32 gets the most visible improvement

The most impactful everyday change is space preallocation in FAT32. Developer Matheus Alves de Almeida moved the work to a helper called \`fat\_preallocate()\` and used a dedicated path to avoid rechecking the file size on every write. In synchronous mounts with small clusters, the published measurements reach improvements of up to 162 times.

A second FAT tweak, merged earlier in the cycle, reached 265 times in the same type of test. These are figures that only appear in specific scenarios, and they should be read that way: they do not describe a normal desktop running ext4, but cards and devices with FAT where every synchronous write came at a premium.

For anyone preparing images for Raspberry Pi, cameras or devices that boot from FAT, the difference shows up in copy times. For everyone else, it is a reminder that the ecosystem's oldest file system is still getting attention decades later.

The improvement should not be mistaken for a transformation. FAT32 still lacks POSIX permissions, symbolic links and journaling, so the gain does not make it a candidate for a file server or for a laptop running ext4 or btrfs. What changes is the cost of a specific operation, not the nature of the system. The previous cycle had already moved [FAT32 with improvements of up to 4.06 times](https://observatorioblockchain.com/en/technology/linux-7-5-fat32-4-06x-airdrop-t2/), and the [ext4 change in 7.4](https://observatorioblockchain.com/en/technology/ext4-iomap-linux-7-4-escrituras/) shares the same philosophy: taking work off the hot path.

## Memory, BPF and the AI trail

The memory management batch arrived through a new channel. David Hildenbrand, an Arm engineer, began routing some of the patches through his own Core MM tree instead of relying solely on Andrew Morton, and that batch included bugs that had been hidden for years.

The scope is broad: transparent huge pages, GUP, KSM, USERFAULTFD, shmem and memfd, and hugetlb. The fixes cover information leaks, out-of-bounds writes, data loss and corruption, use-after-free, deadlocks and loops that cannot be killed. Hildenbrand was explicit about the origin: most of the problems were reported by Anthropic.

That detail connects to the other headline of the cycle. In BPF, Alexei Starovoitov signed off on 40 patches hardening the verifier and the execution sandboxes, and said that in his subsystem fixes now arrive faster than new reports. Torvalds picked it up in his usual tone: if it is true, it is a good sign, and he added that it is best not to hold your breath.

The rest of rc2 includes concurrency protection in Bluetooth, rejection of malformed IPsec marks in xfrm, use-after-free fixes in NVMe, work on the KVM guest MMU and a sweep of 22 patches in hwmon, among hundreds of other minor changes.

## What it means for those who do not compile kernels

For most people, nothing immediate. A release candidate is a test snapshot and should only receive fixes, not new features; installing it in production is a laboratory decision. The recommendation remains to stay on the distribution's stable kernel.

What does change is the underlying schedule. If the pace holds, 7.4 could be released on December 20 with seven candidates, or on the 27th with eight. Greg Kroah-Hartman would adopt it as LTS in early 2027, with years of backported patches behind it.

That horizon explains why so many people are watching this cycle. Anyone with new AMD hardware, FAT disks or container workloads has concrete reasons to wait for the stable release instead of chasing every candidate.

## The tide of patches and its bill

Since Anthropic launched its bug-finding model in April, the volume of errors found in mature software has grown to limits teams had not seen. Microsoft fixed 974 bugs in September, after 622 in July and 206 in June; Oracle reached 1,449 in a single month.

The kernel, with more than 40 million lines, absorbs that tide with volunteer maintainers and a merge window that is already approaching 20,000 submissions. The question underlying rc2 is whether the process holds up or whether the bottleneck shifts from finding bugs to reviewing the patches that fix them.

Torvalds's answer is pragmatic: he prefers that AI-generated work arrive during the merge window, when there is room to review it, and not as an avalanche of emergency fixes in the stabilization weeks.

## What remains to be verified

The BPF signal is a data point from one subsystem, not a general trend, and Torvalds himself qualified it in the announcement. The rest of the kernel is still receiving more reports than it can close, and the schedule of seven or eight candidates depends on no serious regression appearing.

For anyone who wants to test, the source is in Torvalds's tree and the tarball at kernel.org; for anyone who just wants a machine that boots, the answer is the same as a month ago: wait for the December stable release and, if it is adopted as LTS, upgrade then with the reassurance of years of maintenance ahead.
