splice: optional SPLICE_F_MOVE on fuse replies - #1
Open
AliRamberg wants to merge 2 commits into
Open
Conversation
The generic pipe-to-file splice path ignores SPLICE_F_MOVE, but /dev/fuse acts on it: fuse_dev_splice_write() in fs/fuse/dev.c sets cs.move_pages, and the kernel then moves each pipe page into the target inode's page cache instead of copying into it. Only readahead reads are eligible, as fuse_send_readpages() is the only caller that sets page_replace, and a page that cannot be stolen falls back to the copy, so this is a hint. Nothing is negotiated at INIT: the kernel advertises FUSE_SPLICE_MOVE but never reads it back. For a filesystem serving reads by splicing out of a local cache file that turns a duplicate into a migration: the folio leaves the cache file's mapping and becomes the fuse inode's, so the payload is neither copied a second time nor resident twice. Measured on MLPerf Storage unet3d, three runs per arm, kernel 6.8: the cache file holds 1.6 GiB rather than 33.5 GiB, the inode 111.2 GiB rather than 81.2 GiB, and the epoch takes 64.0s rather than 73.0s. A successful move removes the folio from the source file's page cache and waits on its writeback first, so only the filesystem knows when it is sound. It opts in per reply, through ReadResultPipeMove or by implementing SpliceMove() bool on its own ReadResult. The flag has to ride on the result type itself: wrapping a ReadResult hides Stateful() from trySplice, which then copies the payload through userspace instead of splicing it. Pair.WriteTo keeps its signature and flags=0 behaviour; the flag goes through a new Pair.WriteToFlags. Change-Id: Ice312bbcd81eb631a599d605ee0e72a1f9b509ff
arielshaqed
requested changes
Aug 16, 2026
arielshaqed
left a comment
arielshaqed
left a comment
There was a problem hiding this comment.
Excitingly great change! But I think as-is it is a bit brittle. Suggestions inline.
| req.serializeHeader(payloadLen) | ||
|
|
||
| return r.trySplice(req, ReadResultPipe(pair, payloadLen)) | ||
| return r.trySplice(req, &pipeReadResult{pair: pair, size: payloadLen, move: spliceFlags != 0}) |
There was a problem hiding this comment.
- I would have expected
spliceFlags & unix.splice_F_MOVE != 0- in future if we add more bits the existing code will break. - Can we use ReadResultPipeMove?!
| // The filesystem decides per reply whether its pages may be moved. | ||
| spliceFlags := 0 | ||
| if m, ok := readResult.(movableResult); ok && m.SpliceMove() { | ||
| spliceFlags = unix.SPLICE_F_MOVE |
There was a problem hiding this comment.
In future we might add more splice flags, so this is brittle.
Suggested change
| spliceFlags = unix.SPLICE_F_MOVE | |
| spliceFlags |= unix.SPLICE_F_MOVE |
AliRamberg
force-pushed
the
splice-f-move
branch
2 times, most recently
from
August 16, 2026 14:40
38ebe59 to
82af371
Compare
AliRamberg
force-pushed
the
splice-f-move
branch
4 times, most recently
from
August 16, 2026 15:00
89fb90f to
d628bbc
Compare
arielshaqed
approved these changes
Aug 16, 2026
arielshaqed
left a comment
arielshaqed
left a comment
There was a problem hiding this comment.
Great, thanks! Excited about the performance improvements that this unlocks.
Comment on lines
+45
to
+47
| // flaggedResult is a ReadResult carrying splice(2) flags for the write to | ||
| // /dev/fuse. | ||
| type flaggedResult interface { |
There was a problem hiding this comment.
Suggested change
| // flaggedResult is a ReadResult carrying splice(2) flags for the write to | |
| // /dev/fuse. | |
| type flaggedResult interface { | |
| // spliceFlaggedResult is a ReadResultPipe carrying splice(2) flags for the write to | |
| // /dev/fuse. | |
| type spliceFlaggedResult interface { |
Author
|
Absolutely! Bare in mind these numbers are also based on the O_DIRECT flow |
AliRamberg
force-pushed
the
splice-f-move
branch
3 times, most recently
from
August 18, 2026 11:57
a84db19 to
3159c25
Compare
Review fixes on top of the SPLICE_F_MOVE commit: - pipeReadResult stores the splice(2) flags themselves instead of a move bool, so trySplice no longer converts int -> bool -> int across the short-read fixup. The flags go straight to Pair.WriteToFlags. - ReadResultPipeMove becomes ReadResultPipeFlags(pipe, size, flags), and the opt-in interface becomes SpliceFlags() int; adding a second flag needs no new constructor. - The recursive trySplice call is not a retry: the payload already sits in the pipe, so the second pass splices pipe->pipe behind the corrected header rather than reading the file again. Signed-off-by: Yahli Ramberg <lryahli@gmail.com>
AliRamberg
force-pushed
the
splice-f-move
branch
from
August 18, 2026 12:03
3159c25 to
2a66f63
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
/dev/fuseacts onSPLICE_F_MOVEeven though the generic pipe-to-file splice path ignores it:fuse_dev_splice_write()setscs.move_pages(fs/fuse/dev.c:2038), and the kernel then moves each pipe page into the target inode's page cache instead of copying into it.For a filesystem that serves reads by splicing out of a local cache file, that turns a duplicate into a migration — the page leaves the cache file's mapping and becomes the fuse inode's, so the payload is neither copied a second time nor resident twice.
Measured on MLPerf Storage unet3d against an everest read-only mount, 1200 files / 162 GiB, 2 accelerators, kernel 6.8, three interleaved runs per arm:
fadviseSPLICE_F_MOVEWarm -12.3%, device reads -13%, and the duplicate is gone rather than reclaimed:
fadviserecovers two thirds of it and pays a syscall per read, a move costs nothing and keeps the byte.API
Opt-in per reply. A successful move removes the page from the source file's page cache and waits on its writeback first, so only the filesystem knows when it is sound:
A
ReadResultcan also opt in by implementingSpliceMove() bool.The flag rides on the result type itself, not a wrapper: wrapping a
ReadResulthidesStateful()fromtrySplice, which then falls back to copying the payload through userspace — silently, sinceerrRecoverSpliceis not logged.TestReadResultPipeSpliceMovefails on exactly that mistake.Pair.WriteTokeeps its signature andflags=0behaviour; the flag goes through a newPair.WriteToFlags.Limits
Only readahead reads are eligible —
fuse_send_readpages()is the only caller that setspage_replace, anddev.c:1919clearsmove_pageswithout it. Whole aligned pages only (dev.c:951), and any page that cannot be stolen falls back to the copy. Nothing is negotiated at INIT: the kernel advertisesFUSE_SPLICE_MOVEand never reads it back.Verification
gofmt,GOOS=linux/darwin/freebsdbuilds,go vet ./fuse/ ./splice/, andTestReadResultPipeSpliceMoveon linux.Upstream review: https://review.gerrithub.io/c/hanwen/go-fuse/+/1245412 — hanwen's comment on patch set 3 was that a
MountOptionsflag is all-or-nothing. This branch is that rework (per-reply, mount option deleted) and has not been pushed to Gerrit yet.Everest consumes it via
ReadResultPipeMoveon read-only mounts; that branch needs ago.modbump before it can merge.