Introduction and overview

Bcachefs is a modern, general purpose, copy on write filesystem descended from bcache, a block layer cache.

The internal architecture is very different from most existing filesystems where the inode is central and many data structures hang off of the inode. Instead, bcachefs is architected more like a filesystem on top of a relational database, with tables for the different filesystem data types - extents, inodes, dirents, xattrs, et cetera.

bcachefs supports almost all of the same features as other modern COW filesystems, such as ZFS and btrfs, but in general with a cleaner, simpler, higher performance design.

Performance overview

The core of the architecture is a very high performance and very low latency b+ tree, which also is not a conventional b+ tree but more of hybrid, taking concepts from compacting data structures: btree nodes are very large, log structured, and compacted (resorted) as necessary in memory. This means our b+ trees are very shallow compared to other filesystems.

What this means for the end user is that since we require very few seeks or disk reads, filesystem latency is extremely good - especially cache cold filesystem latency, which does not show up in most benchmarks but has a huge impact on real world performance, as well as how fast the system "feels" in normal interactive usage. Latency has been a major focus throughout the codebase - notably, we have assertions that we never hold b+ tree locks while doing IO, and the btree transaction layer makes it easily to aggressively drop and retake locks as needed - one major goal of bcachefs is to be the first general purpose soft realtime filesystem.

Additionally, unlike other COW btrees, btree updates are journalled. This greatly improves our write efficiency on random update workloads, as it means btree writes are only done when we have a large block of updates, or when required by memory reclaim or journal reclaim.

Bucket based allocation

As mentioned bcachefs is descended from bcache, where the ability to efficiently invalidate cached data and reuse disk space was a core design requirement. To make this possible the allocator divides the disk up into buckets, typically 512k to 2M but possibly larger or smaller. Buckets and data pointers have generation numbers: we can reuse a bucket with cached data in it without finding and deleting all the data pointers by incrementing the generation number.

In keeping with the copy-on-write theme of avoiding update in place wherever possible, we never rewrite or overwrite data within a bucket - when we allocate a bucket, we write to it sequentially and then we don’t write to it again until the bucket has been invalidated and the generation number incremented.

This means we require a copying garbage collector to deal with internal fragmentation, when patterns of random writes leave us with many buckets that are partially empty (because the data they contained was overwritten) - copy GC evacuates buckets that are mostly empty by writing the data they contain to new buckets. This also means that we need to reserve space on the device for the copy GC reserve when formatting - typically 8% or 12%.

There are some advantages to structuring the allocator this way, besides being able to support cached data:

• By maintaining multiple write points that are writing to different buckets, we’re able to easily and naturally segregate unrelated IO from different processes, which helps greatly with fragmentation.

• The fast path of the allocator is essentially a simple bump allocator - the disk space allocation is extremely fast

• Fragmentation is generally a non issue unless copygc has to kick in, and it usually doesn’t under typical usage patterns. The allocator and copygc are doing essentially the same things as the flash translation layer in SSDs, but within the filesystem we have much greater visibility into where writes are coming from and how to segregate them, as well as which data is actually live - performance is generally more predictable than with SSDs under similar usage patterns.

• The same algorithms will in the future be used for managing SMR hard drives directly, avoiding the translation layer in the hard drive - doing this work within the filesystem should give much better performance and much more predictable latency.

Feature overview

IO path options

Most options that control the IO path can be set at either the filesystem level or on individual inodes (files and directories). When set on a directory via the bcachefs attr command, they will be automatically applied recursively.

Checksumming

bcachefs supports both metadata and data checksumming - crc32c by default, but stronger checksums are available as well. Enabling data checksumming incurs some performance overhead - besides the checksum calculation, writes have to be bounced for checksum stability (Linux generally cannot guarantee that the buffer being written is not modified in flight), but reads generally do not have to be bounced.

Checksum granularity in bcachefs is at the level of individual extents, which results in smaller metadata but means we have to read entire extents in order to verify the checksum. By default, checksummed and compressed extents are capped at 64k. For most applications and usage scenarios this is an ideal trade off, but small random O_DIRECT reads will incur significant overhead. In the future, checksum granularity will be a per-inode option.

Encryption

bcachefs supports authenticated (AEAD style) encryption - ChaCha20/Poly1305. When encryption is enabled, the poly1305 MAC replaces the normal data and metadata checksums. This style of encryption is superior to typical block layer or filesystem level encryption (usually AES-XTS), which only operates on blocks and doesn’t have a way to store nonces or MACs. In contrast, we store a nonce and cryptographic MAC alongside data pointers, meaning we have a chain of trust up to the superblock (or journal, in the case of unclean shutdowns) and can definitely tell if metadata has been modified, dropped, or replaced with an earlier version. Therefore, replay attacks are not possible, with the exception of an offline rollback of the entire filesystem to a previous version (but see the WARNING below).

Encryption can only be specified for the entire filesystem, not per file or directory - this is because metadata blocks do not belong to a particular file. All data and metadata except for the superblock is encrypted, and all data and metadata is authenticated.

In the future we’ll probably add AES-GCM for platforms that have hardware acceleration for AES, but in the meantime software implementations of ChaCha20 are also quite fast on most platforms.

scrypt is currently used for the key derivation function (KDF), which converts the user supplied passphrase to an encryption key. This is the same function used by Tarsnap and Qubes OS’s backup support. The key derivation is implemented entirely in user-space, so other means of deriving a key can be used in the future without any kernel changes.

To format a filesystem with encryption, use

bcachefs format --encrypted /dev/sda1

You will be prompted for a passphrase. Then, to use an encrypted filesystem use the command

bcachefs unlock /dev/sda1

You will be prompted for the passphrase and the encryption key will be added to your in-kernel keyring; mount, fsck and other commands will then work as usual.

The passphrase on an existing encrypted filesystem can be changed with the bcachefs set-passphrase command. To permanently unlock an encrypted filesystem, use the bcachefs remove-passphrase command - this can be useful when dumping filesystem metadata for debugging by the developers.

There is a wide_macs option which controls the size of the cryptographic MACs stored on disk. By default, only 80 bits are stored, which should be sufficient security for most applications. With the wide_macs option enabled we store the full 128 bit MAC, at the cost of making extents 8 bytes bigger. wide_macs is recommended for cases where an attacker can make repeated attempts at forging a MAC, such as scenarios where the storage device itself is untrusted (but see below).

For technical reasons, bcachefs encryption is unsafe if the underlying storage is snapshotted and rolled back to an earlier version. (Using bcachefs’s own snapshot functionality is safe.) Therefore, one must exercise care when using bcachefs encryption with “fancy” storage devices. It is safe to rely on bcachefs encryption if both of the following hold:

• You trust your drives to not be actively malicious. For the internal storage on your laptop or desktop, this is probably a safe assumption, and if it is not, you likely have much worse problems. However, it is not necessarily a safe assumption for e.g. USB drives or network storage. In those cases you will need to decide for yourself if you are worried about this.

• You are not using “fancy” storage systems that support snapshots. This includes e.g. LVM, ZFS, and loop devices on reflinked or snapshotted files. Most network storage and/or virtualization solutions also support snapshots.

If you are using snapshots, you must make sure that you never mount a snapshotted, encrypted volume, except with -o nochanges. If this rule is violated, an attacker might be able to recover sensitive data that the encryption was supposed to protect . Future versions of bcachefs will not have this limitation. In the meantime, one can make this problem much more difficult to exploit by encrypting the volumes on which bcachefs resides using LUKS, provided that LUKS is above anything that could take a snapshot. For instance, if you are using bcachefs on LVM and might take an LVM snapshot, LUKS would need to be between LVM and bcachefs.

Compression

bcachefs supports gzip, lz4 and zstd compression. As with data checksumming, we compress entire extents, not individual disk blocks - this gives us better compression ratios than other filesystems, at the cost of reduced small random read performance.

Data can also be compressed or recompressed with a different algorithm in the background by the rebalance thread, if the background_compression option is set.

Multiple devices

bcachefs is a multi-device filesystem. Devices need not be the same size: by default, the allocator will stripe across all available devices but biasing in favor of the devices with more free space, so that all devices in the filesystem fill up at the same rate. Devices need not have the same performance characteristics: we track device IO latency and direct reads to the device that is currently fastest.

Replication

bcachefs supports standard RAID1/10 style redundancy with the data_replicas and metadata_replicas options. Layout is not fixed as with RAID10: a given extent can be replicated across any set of devices; the bcachefs fs usage command shows how data is replicated within a filesystem.

Erasure coding

bcachefs also supports Reed-Solomon erasure coding - the same algorithm used by most RAID5/6 implementations) When enabled with the ec option, the desired redundancy is taken from the data_replicas option - erasure coding of metadata is not supported.

Erasure coding works significantly differently from both conventional RAID implementations and other filesystems with similar features. In conventional RAID, the "write hole" is a significant problem - doing a small write within a stripe requires the P and Q (recovery) blocks to be updated as well, and since those writes cannot be done atomically there is a window where the P and Q blocks are inconsistent - meaning that if the system crashes and recovers with a drive missing, reconstruct reads for unrelated data within that stripe will be corrupted.

ZFS avoids this by fragmenting individual writes so that every write becomes a new stripe - this works, but the fragmentation has a negative effect on performance: metadata becomes bigger, and both read and write requests are excessively fragmented. Btrfs’s erasure coding implementation is more conventional, and still subject to the write hole problem.

bcachefs’s erasure coding takes advantage of our copy on write nature - since updating stripes in place is a problem, we simply don’t do that. And since excessively small stripes is a problem for fragmentation, we don’t erasure code individual extents, we erasure code entire buckets - taking advantage of bucket based allocation and copying garbage collection.

When erasure coding is enabled, writes are initially replicated, but one of the replicas is allocated from a bucket that is queued up to be part of a new stripe. When we finish filling up the new stripe, we write out the P and Q buckets and then drop the extra replicas for all the data within that stripe - the effect is similar to full data journalling, and it means that after erasure coding is done the layout of our data on disk is ideal.

Since disks have write caches that are only flushed when we issue a cache flush command - which we only do on journal commit - if we can tweak the allocator so that the buckets used for the extra replicas are reused (and then overwritten again) immediately, this full data journalling should have negligible overhead - this optimization is not implemented yet, however.

Device labels and targets

By default, writes are striped across all devices in a filesystem, but they may be directed to a specific device or set of devices with the various target options. The allocator only prefers to allocate from devices matching the specified target; if those devices are full, it will fall back to allocating from any device in the filesystem.

Target options may refer to a device directly, e.g. foreground_target=/dev/sda1, or they may refer to a device label. A device label is a path delimited by periods - e.g. ssd.ssd1 (and labels need not be unique). This gives us ways of referring to multiple devices in target options: If we specify ssd in a target option, that will refer to all devices with the label ssd or labels that start with ssd. (e.g. ssd.ssd1, ssd.ssd2).

Four target options exist. These options all may be set at the filesystem level (at format time, at mount time, or at runtime via sysfs), or on a particular file or directory:

Caching

When an extent has multiple copies on different devices, some of those copies may be marked as cached. Buckets containing only cached data are discarded as needed by the allocator in LRU order.

When data is moved from one device to another according to the
background_target option, the original copy is left in place but marked as cached. With the promote_target option, the original copy is left unchanged and the new copy on the promote_target device is marked as cached.

To do writeback caching, set foreground_target and promote_target to the cache device, and background_target to the backing device. To do writearound caching, set foreground_target to the backing device and promote_target to the cache device.

Durability

Some devices may be considered to be more reliable than others. For example, we might have a filesystem composed of a hardware RAID array and several NVME flash devices, to be used as cache. We can set replicas=2 so that losing any of the NVME flash devices will not cause us to lose data, and then additionally we can set durability=2 for the hardware RAID device to tell bcachefs that we don’t need extra replicas for data on that device - data on that device will count as two replicas, not just one.

The durability option can also be used for writethrough caching: by setting durability=0 for a device, it can be used as a cache and only as a cache - bcachefs won’t consider copies on that device to count towards the number of replicas we’re supposed to keep.

Reflink

bcachefs supports reflink, similarly to other filesystems with the same feature. cp –reflink will create a copy that shares the underlying storage. Reading from that file will become slightly slower - the extent pointing to that data is moved to the reflink btree (with a refcount added) and in the extents btree we leave a key that points to the indirect extent in the reflink btree, meaning that we now have to do two btree lookups to read from that data instead of just one.

Inline data extents

bcachefs supports inline data extents, controlled by the inline_data option (on by default). When the end of a file is being written and is smaller than half of the filesystem blocksize, it will be written as an inline data extent. Inline data extents can also be reflinked (moved to the reflink btree with a refcount added): as a todo item we also intend to support compressed inline data extents.

Subvolumes and snapshots

bcachefs supports subvolumes and snapshots with a similar userspace interface as btrfs. A new subvolume may be created empty, or it may be created as a snapshot of another subvolume. Snapshots are writeable and may be snapshotted again, creating a tree of snapshots.

Snapshots are very cheap to create: they’re not based on cloning of COW btrees as with btrfs, but instead are based on versioning of individual keys in the btrees. Many thousands or millions of snapshots can be created, with the only limitation being disk space.

The following subcommands exist for managing subvolumes and snapshots:

• bcachefs subvolume create: Create a new, empty subvolume

• bcachefs subvolume destroy: Delete an existing subvolume or snapshot

• bcachefs subvolume snapshot: Create a snapshot of an existing subvolume

A subvolume can also be deleting with a normal rmdir after deleting all the contents, as with rm -rf. Still to be implemented: read-only snapshots, recursive snapshot creation, and a method for recursively listing subvolumes.

Quotas

bcachefs supports conventional user/group/project quotas. Quotas do not currently apply to snapshot subvolumes, because if a file changes ownership in the snapshot it would be ambiguous as to what quota data within that file should be charged to.

When a directory has a project ID set it is inherited automatically by descendants on creation and rename. When renaming a directory would cause the project ID to change we return -EXDEV so that the move is done file by file, so that the project ID is propagated correctly to descendants - thus, project quotas can be used as subdirectory quotas.

Management

Formatting

To format a new bcachefs filesystem use the subcommand bcachefs format, or mkfs.bcachefs. All persistent filesystem-wide options can be specified at format time. For an example of a multi device filesystem with compression, encryption, replication and writeback caching:

bcachefs format --compression=lz4               \
                --encrypted                     \
                --replicas=2                    \
                --label=ssd.ssd1 /dev/sda       \
                --label=ssd.ssd2 /dev/sdb       \
                --label=hdd.hdd1 /dev/sdc       \
                --label=hdd.hdd2 /dev/sdd       \
                --label=hdd.hdd3 /dev/sde       \
                --label=hdd.hdd4 /dev/sdf       \
                --foreground_target=ssd         \
                --promote_target=ssd            \
                --background_target=hdd

Mounting

To mount a multi device filesystem, there are two options. You can specify all component devices, separated by colons, e.g.

mount -t bcachefs /dev/sda:/dev/sdb:/dev/sdc /mnt

Or, use the mount.bcachefs tool to mount by filesystem UUID. Still todo: improve the mount.bcachefs tool to support mounting by filesystem label.

No special handling is needed for recovering from unclean shutdown. Journal replay happens automatically, and diagnostic messages in the dmesg log will indicate whether recovery was from clean or unclean shutdown.

The -o degraded option will allow a filesystem to be mounted without all the devices, but will fail if data would be missing. The -o very_degraded can be used to attempt mounting when data would be missing.

Also relevant is the -o nochanges option. It disallows any and all writes to the underlying devices, pinning dirty data in memory as necessary if for example journal replay was necessary - think of it as a "super read-only" mode. It can be used for data recovery, and for testing version upgrades.

The -o verbose enables additional log output during the mount process.

Fsck

It is possible to run fsck either in userspace with the bcachefs fsck subcommand (also available as fsck.bcachefs, or in the kernel while mounting by specifying the -o fsck mount option. In either case the exact same fsck implementation is being run, only the environment is different. Running fsck in the kernel at mount time has the advantage of somewhat better performance, while running in userspace has the ability to be stopped with ctrl-c and can prompt the user for fixing errors. To fix errors while running fsck in the kernel, use the -o fix_errors option.

The -n option passed to fsck implies the -o nochanges option; bcachefs fsck -ny can be used to test filesystem repair in dry-run mode.

Status of data

The bcachefs fs usage may be used to display filesystem usage broken out in various ways. Data usage is broken out by type: superblock, journal, btree, data, cached data, and parity, and by which sets of devices extents are replicated across. We also give per-device usage which includes fragmentation due to partially used buckets.

Journal

The journal has a number of tunables that affect filesystem performance. Journal commits are fairly expensive operations as they require issuing FLUSH and FUA operations to the underlying devices. By default, we issue a journal flush one second after a filesystem update has been done; this is controlled with the journal_flush_delay option, which takes a parameter in milliseconds.

Filesystem sync and fsync operations issue journal flushes; this can be disabled with the journal_flush_disabled option - the journal_flush_delay option will still apply, and in the event of a system crash we will never lose more than (by default) one second of work. This option may be useful on a personal workstation or laptop, and perhaps less appropriate on a server.

The journal reclaim thread runs in the background, kicking off btree node writes and btree key cache flushes to free up space in the journal. Even in the absence of space pressure it will run slowly in the background: this is controlled by the journal_reclaim_delay parameter, with a default of 100 milliseconds.

The journal should be sized sufficiently that bursts of activity do not fill up the journal too quickly; also, a larger journal mean that we can queue up larger btree writes. The bcachefs device resize-journal can be used for resizing the journal on disk on a particular device - it can be used on a mounted or unmounted filesystem.

In the future, we should implement a method to see how much space is currently utilized in the journal.

Device management

Filesystem resize

A filesystem can be resized on a particular device with the bcachefs device resize subcommand. Currently only growing is supported, not shrinking.

Device add/removal

The following subcommands exist for adding and removing devices from a mounted filesystem:

• bcachefs device add: Formats and adds a new device to an existing filesystem.

• bcachefs device remove: Permenantly removes a device from an existing filesystem.

• bcachefs device online: Connects a device to a running filesystem that was mounted without it (i.e. in degraded mode)

• bcachefs device offline: Disconnects a device from a mounted filesystem without removing it.

• bcachefs device evacuate: Migrates data off of a particular device to prepare for removal, setting it read-only if necessary.

• bcachefs device set-state: Changes the state of a member device: one of rw (readwrite), ro (readonly), failed, or spare.

  A failed device is considered to have 0 durability, and replicas on that device won’t be counted towards the number of replicas an extent should have by rereplicate - however, bcachefs will still attempt to read from devices marked as failed.

The bcachefs device remove, bcachefs device offline and bcachefs device set-state commands take force options for when they would leave the filesystem degraded or with data missing. Todo: regularize and improve those options.

Data management

Data rereplicate

The bcachefs data rereplicate command may be used to scan for extents that have insufficient replicas and write additional replicas, e.g. after a device has been removed from a filesystem or after replication has been enabled or increased.

Rebalance

To be implemented: a command for moving data between devices to equalize usage on each device. Not normally required because the allocator attempts to equalize usage across devices as it stripes, but can be necessary in certain scenarios - i.e. when a two-device filesystem with replication enabled that is very full has a third device added.

Scrub

To be implemented: a command for reading all data within a filesystem and ensuring that checksums are valid, fixing bitrot when a valid copy can be found.

Options

Most bcachefs options can be set filesystem wide, and a significant subset can also be set on inodes (files and directories), overriding the global defaults. Filesystem wide options may be set when formatting, when mounting, or at runtime via /sys/fs/bcachefs/<uuid>/options/. When set at runtime via sysfs the persistent options in the superblock are updated as well; when options are passed as mount parameters the persistent options are unmodified.

File and directory options

<say something here about how attrs must be set via bcachefs attr command>

Options set on inodes (files and directories) are automatically inherited by their descendants, and inodes also record whether a given option was explicitly set or inherited from their parent. When renaming a directory would cause inherited attributes to change we fail the rename with -EXDEV, causing userspace to do the rename file by file so that inherited attributes stay consistent.

Inode options are available as extended attributes. The options that have been explicitly set are available under the bcachefs namespace, and the effective options (explicitly set and inherited options) are available under the bcachefs_effective namespace. Examples of listing options with the getfattr command:

$ getfattr -d -m '^bcachefs\.' filename
$ getfattr -d -m '^bcachefs_effective\.' filename

Options may be set via the extended attribute interface, but it is preferable to use the bcachefs setattr command as it will correctly propagate options recursively.

Full option list

Error actions

The errors option is used for inconsistencies that indicate some sort of a bug. Valid error actions are:

continue

Log the error but continue normal operation

ro

Emergency read only, immediately halting any changes to the filesystem on disk

panic

Immediately halt the entire machine, printing a backtrace on the system console

Checksum types

Valid checksum types are:

none

crc32c

(default)

crc64

Compression types

Valid compression types are:

none

(default)

lz4

gzip

zstd

String hash types

Valid hash types for string hash tables are:

crc32c

crc64

siphash

(default)

Debugging tools

Sysfs interface

Mounted filesystems are available in sysfs at /sys/fs/bcachefs/<uuid>/ with various options, performance counters and internal debugging aids.

Options

Filesystem options may be viewed and changed via
/sys/fs/bcachefs/<uuid>/options/, and settings changed via sysfs will be persistently changed in the superblock as well.

Time stats

bcachefs tracks the latency and frequency of various operations and events, with quantiles for latency/duration in the /sys/fs/bcachefs/<uuid>/time_stats/ directory.

Internals

Unit and performance tests

Echoing into /sys/fs/bcachefs/<uuid>/perf_test runs various low level btree tests, some intended as unit tests and others as performance tests. The syntax is

    echo <test_name> <nr_iterations> <nr_threads> > perf_test

When complete, the elapsed time will be printed in the dmesg log. The full list of tests that can be run can be found near the bottom of fs/bcachefs/tests.c.

Debugfs interface

The contents of every btree, as well as various internal per-btree-node information, are available under /sys/kernel/debug/bcachefs/<uuid>/.

For every btree, we have the following files:

Listing and dumping filesystem metadata

bcachefs show-super

This subcommand is used for examining and printing bcachefs superblocks. It takes two optional parameters:

bcachefs list

This subcommand gives access to the same functionality as the debugfs interface, listing btree nodes and contents, but for offline filesystems.

bcachefs list_journal

This subcommand lists the contents of the journal, which primarily records btree updates ordered by when they occured.

bcachefs dump

This subcommand can dump all metadata in a filesystem (including multi device filesystems) as qcow2 images: when encountering issues that fsck can not recover from and need attention from the developers, this makes it possible to send the developers only the required metadata. Encrypted filesystems must first be unlocked with bcachefs remove-passphrase.

ioctl interface

This section documents bcachefs-specific ioctls:

On disk format

Superblock

The superblock is the first thing to be read when accessing a bcachefs filesystem. It is located 4kb from the start of the device, with redundant copies elsewhere - typically one immediately after the first superblock, and one at the end of the device.

The bch_sb_layout records the amount of space reserved for the superblock as well as the locations of all the superblocks. It is included with every superblock, and additionally written 3584 bytes from the start of the device (512 bytes before the first superblock).

Most of the superblock is identical across each device. The exceptions are the dev_idx field, and the journal section which gives the location of the journal.

The main section of the superblock contains UUIDs, version numbers, number of devices within the filesystem and device index, block size, filesystem creation time, and various options and settings. The superblock also has a number of variable length sections:

Journal

Every journal write (struct jset) contains a list of entries: struct jset_entry. Below are listed the various journal entry types.

Btrees

Btree keys