Ext3 was limited to a total of 32,000 subdirectories; ext4 allows an unlimited number. Beginning with kernel 2.6.23, ext4 uses HTree indices to mitigate performance loss with huge numbers of subdirectories.
#### Journal checksumming
Ext3 did not checksum its journals, which presented problems for disk or controller devices with caches of their own, outside the kernel's direct control. If a controller or a disk with its own cache did writes out of order, it could break ext3's journaling transaction order, potentially corrupting files being written to during (or for some time preceding) a crash.
In theory, this problem is resolved by the use of write barriers—when mounting the filesystem, you set `barrier=1` in the mount options, and the device will then honor `fsync()` calls all the way down to the metal. In practice, it's been discovered that storage devices and controllers frequently do not honor write barriers—improving performance (and benchmarks, where they're compared to their competitors) but opening up the possibility of data corruption that should have been prevented.
Checksumming the journal allows the filesystem to realize that some of its entries are invalid or out-of-order on the first mount after a crash. This thereby avoids the mistake of rolling back partial or out-of-order journal entries and further damaging the filesystem—even if the storage devices lie and don't honor barriers.
#### Fast filesystem checks
Under ext3, the entire filesystem—including deleted and empty files—required checking when `fsck` is invoked. By contrast, ext4 marks unallocated blocks and sections of the inode table as such, allowing `fsck` to skip them entirely. This greatly reduces the time to run `fsck` on most filesystems and has been implemented since kernel 2.6.24.
#### Improved timestamps
Ext3 offered timestamps granular to one second. While sufficient for most uses, mission-critical applications are frequently looking for much, much tighter time control. Ext4 makes itself available to those enterprise, scientific, and mission-critical applications by offering timestamps in the nanoseconds.
Ext3 filesystems also did not provide sufficient bits to store dates beyond January 18, 2038. Ext4 adds an additional two bits here, extending [the Unix epoch][5] another 408 years. If you're reading this in 2446 AD, you have hopefully already moved onto a better filesystem—but it'll make me posthumously very, very happy if you're still measuring the time since UTC 00:00, January 1, 1970.
#### Online defragmentation
Neither ext2 nor ext3 directly supported online defragmentation—that is, defragging the filesystem while mounted. Ext2 had an included utility, **e2defrag** , that did what the name implies—but it needed to be run offline while the filesystem was not mounted. (This is, obviously, especially problematic for a root filesystem.) The situation was even worse in ext3—although ext3 was much less likely to suffer from severe fragmentation than ext2 was, running **e2defrag** against an ext3 filesystem could result in catastrophic corruption and data loss.
Although ext3 was originally deemed "unaffected by fragmentation," processes that employ massively parallel write processes to the same file (e.g., BitTorrent) made it clear that this wasn't entirely the case. Several userspace hacks and workarounds, such as [Shake][6], addressed this in one way or another—but they were slower and in various ways less satisfactory than a true, filesystem-aware, kernel-level defrag process.
Ext4 addresses this problem head on with **e4defrag** , an online, kernel-mode, filesystem-aware, block-and-extent-level defragmentation utility.
### Ongoing ext4 development
Ext4 is, as the Monty Python plague victim once said, "not quite dead yet!" Although [its principal developer regards it][7] as a mere stopgap along the way to a truly [next-generation filesystem][8], none of the likely candidates will be ready (due to either technical or licensing problems) for deployment as a root filesystem for some time yet.
There are still a few key features being developed into future versions of ext4, including metadata checksumming, first-class quota support, and large allocation blocks.
#### Metadata checksumming
Since ext4 has redundant superblocks, checksumming the metadata within them offers the filesystem a way to figure out for itself whether the primary superblock is corrupt and needs to use an alternate. It is possible to recover from a corrupt superblock without checksumming—but the user would first need to realize that it was corrupt, and then try manually mounting the filesystem using an alternate. Since mounting a filesystem read-write with a corrupt primary superblock can, in some cases, cause further damage, this isn't a sufficient solution, even with a sufficiently experienced user!
Compared to the extremely robust per-block checksumming offered by next-gen filesystems such as btrfs or zfs, ext4's metadata checksumming is a pretty weak feature. But it's much better than nothing.
Although it sounds like a no-brainer—yes, checksum ALL THE THINGS!—there are some significant challenges to bolting checksums into a filesystem after the fact; see [the design document][9] for the gritty details.
#### First-class quota support
Wait, quotas?! We've had those since the ext2 days! Yes, but they've always been an afterthought, and they've always kinda sucked. It's probably not worth going into the hairy details here, but the [design document][10] lays out the ways quotas will be moved from userspace into the kernel and more correctly and performantly enforced.
#### Large allocation blocks
As time goes by, those pesky storage systems keep getting bigger and bigger. With some solid-state drives already using 8K hardware blocksizes, ext4's current limitation to 4K blocks gets more and more limiting. Larger storage blocks can decrease fragmentation and increase performance significantly, at the cost of increased "slack" space (the space left over when you only need part of a block to store a file or the last piece of a file).
You can view the hairy details in the [design document][11].
### Practical limitations of ext4
Ext4 is a robust, stable filesystem, and it's what most people should probably be using as a root filesystem in 2018. But it can't handle everything. Let's talk briefly about some of the things you shouldn't expect from ext4—now or probably in the future.
Although ext4 can address up to 1 EiB—equivalent to 1,000,000 TiB—of data, you really, really shouldn't try to do so. There are problems of scale above and beyond merely being able to remember the addresses of a lot more blocks, and ext4 does not now (and likely will not ever) scale very well beyond 50-100 TiB of data.
Ext4 also doesn't do enough to guarantee the integrity of your data. As big an advancement as journaling was back in the ext3 days, it does not cover a lot of the common causes of data corruption. If data is [corrupted][12] while already on disk—by faulty hardware, impact of cosmic rays (yes, really), or simple degradation of data over time—ext4 has no way of either detecting or repairing such corruption.
Building on the last two items, ext4 is only a pure filesystem, and not a storage volume manager. This means that even if you've got multiple disks—and therefore parity or redundancy, which you could theoretically recover corrupt data from—ext4 has no way of knowing that or using it to your benefit. While it's theoretically possible to separate a filesystem and storage volume management system in discrete layers without losing automatic corruption detection and repair features, that isn't how current storage systems are designed, and it would present significant challenges to new designs.
### Alternate filesystems
Before we get started, a word of warning: Be very careful with any alternate filesystem which isn't built into and directly supported as a part of your distribution's mainline kernel!
Even if a filesystem is safe, using it as the root filesystem can be absolutely terrifying if something hiccups during a kernel upgrade. If you aren't extremely comfortable with the idea of booting from alternate media and poking manually and patiently at kernel modules, grub configs, and DKMS from a chroot... don't go off the reservation with the root filesystem on a system that matters to you.
There may well be good reasons to use a filesystem your distro doesn't directly support—but if you do, I strongly recommend you mount it after the system is up and usable. (For example, you might have an ext4 root filesystem, but store most of your data on a zfs or btrfs pool.)
#### XFS
XFS is about as mainline as a non-ext filesystem gets under Linux. It's a 64-bit, journaling filesystem that has been built into the Linux kernel since 2001 and offers high performance for large filesystems and high degrees of concurrency (i.e., a really large number of processes all writing to the filesystem at once).
XFS became the default filesystem for Red Hat Enterprise Linux, as of RHEL 7. It still has a few disadvantages for home or small business users—most notably, it's a real pain to resize an existing XFS filesystem, to the point it usually makes more sense to create another one and copy your data over.
While XFS is stable and performant, there's not enough of a concrete end-use difference between it and ext4 to recommend its use anywhere that it isn't the default (e.g., RHEL7) unless it addresses a specific problem you're having with ext4, such as >50 TiB capacity filesystems.
XFS is not in any way a "next-generation" filesystem in the ways that ZFS, btrfs, or even WAFL (a proprietary SAN filesystem) are. Like ext4, it should most likely be considered a stopgap along the way towards [something better][8].
#### ZFS
ZFS was developed by Sun Microsystems and named after the zettabyte—equivalent to 1 trillion gigabytes—as it could theoretically address storage systems that large.
A true next-generation filesystem, ZFS offers volume management (the ability to address multiple individual storage devices in a single filesystem), block-level cryptographic checksumming (allowing detection of data corruption with an extremely high accuracy rate), [automatic corruption repair][12] (where redundant or parity storage is available), rapid [asynchronous incremental replication][13], inline compression, and more. [A lot more][14].
The biggest problem with ZFS, from a Linux user's perspective, is the licensing. ZFS was licensed CDDL, which is a semi-permissive license that conflicts with the GPL. There is a lot of controversy over the implications of using ZFS with the Linux kernel, with opinions ranging from "it's a GPL violation" to "it's a CDDL violation" to "it's perfectly fine, it just hasn't been tested in court." Most notably, Canonical has included ZFS code inline in its default kernels since 2016 without legal challenge so far.
At this time, even as a very avid ZFS user myself, I would not recommend ZFS as a root Linux filesystem. If you want to leverage the benefits of ZFS on Linux, set up a small root filesystem on ext4, then put ZFS on your remaining storage, and put data, applications, whatever you like on it—but keep root on ext4, until your distribution explicitly supports a zfs root.
#### btrfs
Btrfs—short for B-Tree Filesystem, and usually pronounced "butter"—was announced by Chris Mason in 2007 during his tenure at Oracle. Btrfs aims at most of the same goals as ZFS, offering multiple device management, per-block checksumming, asynchronous replication, inline compression, and [more][8].
As of 2018, btrfs is reasonably stable and usable as a standard single-disk filesystem but should probably not be relied on as a volume manager. It suffers from significant performance problems compared to ext4, XFS, or ZFS in many common use cases, and its next-generation features—replication, multiple-disk topologies, and snapshot management—can be pretty buggy, with results ranging from catastrophically reduced performance to actual data loss.
The ongoing status of btrfs is controversial; SUSE Enterprise Linux adopted it as its default filesystem in 2015, whereas Red Hat announced it would no longer support btrfs beginning with RHEL 7.4 in 2017. It is probably worth noting that production, supported deployments of btrfs use it as a single-disk filesystem, not as a multiple-disk volume manager a la ZFS—even Synology, which uses btrfs on its storage appliances, but layers it atop conventional Linux kernel RAID (mdraid) to manage the disks.
