mergerfs/mkdocs/docs/config/passthrough.md

passthrough.io

• default: off
• arguments:
  • off: Passthrough is never enabled.
  • ro: Only enable IO passthrough when file opened for reading only.
  • wo: Only enable IO passthrough when file opened for writing only.
  • rw: Enable IO passthrough when file opened for reading, writing, or both.

In Linux 6.9 a IO passthrough feature was added to FUSE. Typically mergerfs has to act as an active proxy for all read and write requests. This results in, at times, significant overhead compared to direct interaction with the underlying filesystems. Not because mergerfs is doing anything particularly slow but because the additional communication and data transfers required. With the passthrough.io feature mergerfs is able to instruct the kernel to perform reads and writes directly on the underlying file. Bypassing mergerfs entirely (specifically and only for reads and writes) and thereby providing near native performance.

This performance does come at the cost of some functionality as mergerfs no longer has control over reads and writes which means all features related to read/write are affected.

• moveonenospc: Does not work because errors are not reported back to mergerfs.
• nullrw: Since mergerfs no longer receives read/write requests there is no read or write to ignore.
• readahead: Still affects the readahead values of underlying filesystems and mergerfs itself but no longer relevant to mergerfs in that it is not servicing IO requests.
• fuse_msg_size: The primary reason to increase the FUSE message size is to allow transferring more data per request which helps improve read and write performance. Without read/write requests being sent to mergerfs there is little reason to have larger message sizes since the only other message larger than 1 page, directory reading, currently has a small fixed buffer size.
• parallel-direct-writes: Not only is direct-io and passthrough.io mutually exclusive but since mergerfs no longer receives write requests in passthrough mode then there is no parallel writes possible.
• cache.writeback: FUSE's writeback caching is incompatible with passthrough.io. If cache.writeback=true when enabling passthrough.io it will be reset to false.
• cache.files: Must be enabled for passthrough.io to work. When cache.files=off FUSE's direct-io mode is enabled which overrides passthrough.io. Meaning cache.files should be set to partial, full or auto-full to use passthrough.io. If cache.files is set to off when using passthrough.io it will reset it to auto-full.

Technically mergerfs has the ability to choose options like passthrough.io, direct-io, and page caching independently for every file opened. However, at the moment there is no use case for picking and choosing which to enable outside cache.files=per-process (which is largely unnecessary on Linux v6.6 and above. See direct-io-allow-mmap) If such a use case arises please reach out to the author to discuss.

Unlike preload.so, passthrough.io will work for any software interacting with mergerfs. However, passthrough.io requires Linux v6.9 or above to work.

NOTE: This feature will ONLY work if mergerfs is running as root as currently only root is allowed to leverage the kernel feature.

NOTE: If a file has been opened and passthrough.io enabled, while that file is open, if another open request is made mergerfs must also enable passthrough.io for the second open request. This is a limitation of how the passthrough feature works.

NOTE: In order to add passthrough.io feature to mergerfs it was necessary to remove the "feature" where mergerfs could open the same file on different branches. Such as using func.open=rand and having multiple files at the same relative path across different branches. This "feature" was very very rarely used and it was impossible to support passthrough.io without changing that behavior.

Alternatives

• preload.so: Leverages the ability to intercept certain filesystem calls to bypass mergerfs for IO. See page for details and limitations.

Benchmarks

Simple dd based benchmark on a Hades Canyon (Intel i7-8809G, 32GB DDR4-2400 RAM).

# /mnt/test is mergerfs over a single tmpfs mount
dd if=/dev/zero of=/mnt/test/tmpfile bs=1M count=2000 oflag=nocache
conv=fdatasync status=progress

Summary: passthrough is ~95% native speed. 200% faster than direct-io (cache.files=off).

• straight to tmpfs: 1.7 GB/s
• nullrw=true: 2.1 GB/s
• cache.files=off; passthrough=off: 800 MB/s
• cache.files=auto-full; passthrough=rw: 1.6 GB/s
• cache.files=auto-full; cache.writeback=false: 518 MB/s
• cache.files=auto-full; cache.writeback=true: 518 MB/s