# ext4 speeds up large writes in Linux 7.4 thanks to improvements

> EXT4 Linux 7.4 debuts an optional IOmap path with up to 105% more speed on 1 MiB writes, but only if you enable it manually.
> By José Maldonado · Technology · 2026-10-28

Large writes in EXT4 have just stopped taking the same path as in 2006. The most widely deployed filesystem on Linux servers debuts a buffered input/output path among the new features of EXT4 Linux 7.4: it abandons classic block handling and relies on IOmap, the framework already used by XFS and other modern filesystems to translate file positions into disk blocks. The change landed in the main tree on the night of Sunday 25 to Monday 26 October, when Linus Torvalds accepted the pull that Theodore Ts’o had sent on Thursday 22 and about which he had to ask on Sunday.

Behind the new feature are 31 patches signed by Zhang Yi, a Huawei engineer who has spent more than two years pushing this conversion. The tests accompanying the series, measured with FIO on a four-core virtual machine and a memory-backed block device, show a 105% jump in 1 MiB writes that bypass the cache and 30% in 64 KiB writes with background writeback. These are the author's own figures, not those of an independent test bench, and they should be read with that caution.

The fine print explains why the news hasn't reached your desktop. The new path does not enable itself: it requires the \`buffered\_iomap\` mount option, coexists poorly with half a dozen classic EXT4 features and disables itself when it encounters one of them. Anyone administering a NAS or a backup server has reasons to try it this quarter; anyone who just opens a browser and saves documents will barely notice the difference, because there the bottleneck is almost never the filesystem.

## From buffer heads to IOmap, thirty-one patches later

EXT4 was born inheriting EXT3's plumbing, which describes the state of each block in memory with structures called buffer heads. Each one carries its own accounting and its own lock, so writing a multi-gigabyte file means managing hundreds of thousands of small objects.

The [IOmap](https://docs.kernel.org/filesystems/iomap/) framework changes the approach: it describes block ranges and works with folios, the memory pages the kernel has handled for years instead of loose 4 KiB pages.

Zhang Yi did not convert only writes. The series also rewrites buffered reads, writeback, the mmap path and the splitting of ranges into partial blocks, as well as reordering the zeroing of the last block of a growing file. That detail is the one that causes the most headaches, because it touches the update of \`i\_disksize\`, the size EXT4 records in the inode before the file is actually on disk. Ts’o described the result as "significant" improvements in large writes and added assorted fixes for fast commit and to avoid deadlocks and out-of-bounds reads when the filesystem is corrupted.

The merge required waiting. Torvalds had already closed the architecture and driver branches and returned to filesystems on Sunday night, with a wry comment about how slow everything was going because of two days of diving. The [pull thread](https://lore.kernel.org/linux-ext4/atoWDJDx6sFovvjp@mit.edu/T/) leaves the merge sealed at 02:20 UTC on Monday 26.

## What did the author measure with FIO?

The figures in the announcement come from a battery with block sizes from 1 KiB to 1 MiB, with and without fsync, with and without rewriting existing blocks and with the uncached write flag the kernel exposes to applications. The numbers are in MiB per second and compare the classic path with the IOmap one:

Write scenario

EXT4 with buffer heads (MiB/s)

EXT4 with IOmap (MiB/s)

Difference

1 MiB, page cache only

4.295

5.780

+35%

64 KiB, with background writeback

734

957

+30%

64 KiB, uncached

415

698

+68%

1 MiB, uncached

1.404

2.883

+105%

1 KiB with fsync after each write

0,71

0,60

\-15%

The table tells an uncomfortable story for easy marketing. Where the block is large and data flows in bulk, the new path nearly doubles performance. Where the pattern is a tiny write followed by fsync, the result gets 7% to 15% worse, and in reads the difference is not significant. The [full series](https://lwn.net/Articles/1099660/) documents those regressions instead of hiding them, something unusual in a patch that reaches mainline.

This is where the news becomes useful for deciding. A backup server that rewrites virtual machine image files, a NAS that receives copies via rsync or a store of compressed logs work right in the band that improves. A small database with frequent transactions, no.

## The features left parked

The new path does not yet cover everything EXT4 can do, and that gap is why the switch is off by default. If the filesystem needs any of these pieces, the kernel returns the inode to the old path without asking permission:

*   Online defragmentation, the one that reorders a file while it is still open.
*   Inline data, the trick that puts tiny contents inside the inode itself.
*   Integrity verification with fs-verity and per-file encryption with fscrypt.
*   Inodes without extents, an inheritance from volumes created with EXT3.
*   \`data=journal\` mode, which some servers use for databases.

A freshly formatted volume with default values does qualify. An old volume, one encrypted with fscrypt or one converted from EXT3 stays out until the work advances, because each of those features requires touching locking paths the patch does not yet cover.

## The price of turning off data=ordered

There is a second trade-off, less visible in the manual and more relevant to anyone who cares for data. The IOmap path does not use \`data=ordered\` mode, the mechanism that guarantees data reaches disk before the metadata describing it.

Instead, EXT4 reserves unwritten extents for extensions and postpones the update of \`i\_disksize\` until the zeroed block at the end is actually on disk. Zhang Yi himself admits the compromise in the cover letter: after a power cut, the inode size may be updated while the content of that tail reads as zeros.

It is worth noting that the broken guarantee is narrow and that the author considers it acceptable. Even so, anyone who chooses \`data=ordered\` for its behavior after crashes should think twice before trading it for sequential speed. The opposite option exists, it is called \`nobuffered\_iomap\` and it turns the switch off without unmounting anything else.

## Where the improvement shows (and where it doesn't)

The rest of the 7.4 window had accumulated 6,753 changes without merges by Friday 23, according to the [merge window tally](https://observatorioblockchain.com/en/technology/linux-7-4-merge-window-6753-cambios/), and this EXT4 piece is one of those that do not appear on any distribution's front page. No one is going to notice the change when booting the desktop. Whoever copies a 400 GB backup and watches the clock will, or whoever moves disk images between two EXT4 volumes on the same machine.

Nor is it the only new feature of the cycle. Linux 7.3, released on Sunday 18, had already moved pieces in the scheduler and in AMD GPU memory, according to the [summary of those weeks](https://observatorioblockchain.com/en/technology/linux-7-3-failfs-scheduler-gaming-vram/). With 7.4 still in the window, what is arriving is foundations.

## A switch waiting its turn

With the code already in mainline, the question becomes when it stops being optional. Ts’o was explicit that the path stays behind a mount option precisely because features are missing, and the series leaves open notes on what remains to be done: online defragmentation, inline data and compatibility with old volumes. The first release candidate of 7.4 will not arrive before 1 November, so no stable distribution will see it this month.

In the meantime, the practical advice fits in one line of \`/etc/fstab\` and one test: mount a lab volume with \`buffered\_iomap\`, copy a large file onto it, compare times and check whether that volume uses encryption or \`data=journal\`. If any of that appears, the new path will not activate and you will have lost nothing. If the volume is clean, you will gain the performance almost no one measures and almost everyone appreciates when the backup clock approaches dawn.
