feat(cpio): improve initramfs image performance and efficiency via cpio reflinks #1531

ddiss · 2021-06-08T23:55:18Z

Changes

This patchset attempts to speed up initramfs generation for some common (Btrfs / XFS) setups by having Dracut make heavier use of reflinks (AKA copy-on-write clones) during initramfs generation. A good portion of an uncompressed+unstripped initramfs image is duplicate data, which really shouldn't need to be shuffled around when on the same COW clone capable FS.

This is a rework of my #1148 feature submission. Instead for relying on the GNU cpio patchset for copy-on-write FS optimized reflink I/O, a new dracut-cpio tool is provided for cpio archive creation.
My motivating factors for dropping GNU cpio in favour of dracut-cpio are:

GNU cpio appears to receive very little attention upstream
relative simplicity of implementation
- dracut-cpio only needs to provide support for cpio newc archive creation for extraction by the kernel
- rust's standard library supports copy_file_range() natively

The new dracut-cpio functionality is disabled by default. It can be explicitly enabled by building with the --enable-dracut-cpio configure option and then calling dracut with --enhanced-cpio.

Performance

Preliminary benchmarks indicate a significant improvement in initramfs creation and kernel extraction times with reflinks, as shown via the Dracut runtime and QEMU boot time values respectively.
Storage utilization is also improved with initramfs reflinks, as shown by the significant reduction in exclusive extents. Shared extents represent space reclaimed through deduplication of initramfs source file and cpio data segments.

Btrfs	Dracut `xz`	Dracut `zstd`	Dracut reflink
Dracut runtime	16.809s (ref)	12.935s (-23.05%)	10.292s (-38.59%)
QEMU boot time: kernel to /init	2361ms (ref)	951.0ms (-59.71%)	857.8ms (-63.66%)
initramfs `fiemap` data
total bytes	14647296 (ref)	16191488 (+10.54%)	34447360 (+135.2%)
shared (deduplicated)	0 (ref)	0 (+-0.00%)	31309824 (note)
exclusive	14647296 (ref)	16191488 (+10.54%)	3137536 (note)

note: Btrfs fiemap results varied significantly across multiple runs, therefore shared and exclusive values above should only be seen as a guide. In contrast, XFS results were consistent across all runs:

XFS	Dracut `xz`	Dracut `zstd`	Dracut reflink
Dracut runtime	14.180s (ref)	10.437s (-26.40%)	9.7687s (-31.11%)
QEMU boot time: kernel to /init	2373ms (ref)	950.9ms (-59.92%)	859.5ms (-63.78%)
initramfs `fiemap` data
total bytes	14675968 (ref)	16158720 (+10.10%)	34689024 (+136.4%)
shared (deduplicated)	0 (ref)	0 (+-0.00%)	26525696
exclusive	14675968 (ref)	16158720 (+10.10%)	8163328 (-44.37%)

The Dracut xz case corresponds to SUSE Dracut version 055, patched with dracut-cpio reflink support, but configured to run with current SUSE defaults:
GNU cpio archiving, xz -0 --check=crc32 --memlimit-compress=50% compression and strip symbol discard.
The Dracut zstd case matches Dracut xz except for compression which is configured to use zstd -3 -T0.
The Dracut reflink case has dracut-cpio alignment and reflinks enabled via enhanced_cpio=yes. To ensure successful extent sharing, initramfs compression and symbol discard are disabled, alongside use of a reflink friendly Dracut staging area (tmpdir=/boot).

These benchmarks were performed on Tumbleweed 20210924 (5.14.6 kernel) virtual machines assigned 2 vCPUs and 8GiB RAM. The QEMU/KVM hypervisor host was running the same OS and stored the raw VM disk images in memory (tmpfs).
Importantly, bootloader initramfs image read times were not evaluated due to a lack of time and existing instrumentation.

Memory backed storage is significantly less vulnerable to the fragmentation effects of CoW reflinks compared to HDDs. Further benchmarking on SSDs and HDDs is necessary, at least before considering dracut-cpio as a complete replacement for GNU cpio archive creation. The benchmarks were done on a completely different machine and kernel to #1148 , so shouldn't be used for comparison with those numbers.

Special thanks to @Firstyear for helping me get my rust changes in shape for upstream submission.

Checklist

I have tested it locally
I have reviewed and updated any documentation if relevant
I am providing new code and test(s) for it

rapid-cpio/Cargo.toml

configure

man/dracut.8.asc

dracut.sh

johannbg · 2021-06-12T06:55:02Z

@ddiss note that @haraldh is on vacation until end of June so expect late review from him

stale · 2021-07-12T07:05:57Z

This issue is being marked as stale because it has not had any recent activity. It will be closed if no further activity occurs. If this is still an issue in the latest release of Dracut and you would like to keep it open please comment on this issue within the next 7 days. Thank you for your contributions.

johannbg · 2021-07-17T08:34:30Z

@ddiss hows the progress going on @Firstyear and @tpgxyz pointers ? @haraldh ping

ddiss · 2021-07-26T11:42:04Z

@ddiss hows the progress going on @Firstyear and @tpgxyz pointers ? @haraldh ping

I'm back from leave so will return to this in the coming days. Thanks for the ping.

the8472 · 2021-07-29T21:57:53Z

rust's standard library supports copy_file_range() natively

Please keep in mind that in std it is considered an optimization and not part of our API guarantees. We had to disable copy_file_range before due to regressions. So if you want to rely on that it you may want to write a test to have a canary. E.g. by tracing syscalls for a particular test or (assuming a suitable filesystem) checking that extents really are shared.

Other options would be using a crate that makes stronger guarantees or documenting that it's best-effort.

ddiss · 2021-07-29T22:05:59Z

rust's standard library supports copy_file_range() natively

Please keep in mind that in std it is considered an optimization and not part of our API guarantees. We had to disable copy_file_range before due to regressions. So if you want to rely on that it you may want to write a test to have a canary. E.g. by tracing syscalls for a particular test or (assuming a suitable filesystem) checking that extents really are shared.

Other options would be using a crate that makes stronger guarantees or documenting that it's best-effort.

Ack, understood. Best-effort is fine, as that's what we already get from the kernel with regard to performing COW reflink or splice fallback. Thanks for the clarification.

ddiss · 2021-08-19T13:27:56Z

Changes since last version:

perform configure time check for Cargo presence
fix lint-shell warnings

ddiss · 2021-08-27T16:36:06Z

Changes since last version:

more shellcheck fixes
move dracut-cpio sources under src/dracut-cpio

TomasTomecek · 2021-08-28T07:50:18Z

/packit build

ddiss · 2021-09-02T16:08:15Z

Changes since last version:

rename --cpio-reflink parameter to --enhanced-cpio to reflect the fact that it won't necessarily result in reflinked data
support compression alongside dracut-cpio
move dinfo call after dracut-init.sh is loaded (thanks @mwilck)
fix cmdline vs config parameter precedence (thanks @mwilck)
minor debug output and variable rewording

ddiss · 2021-09-03T16:31:30Z

I've tacked on one extra commit which adds test coverage for the dracut.sh --enhanced-cpio code-path, as well as kernel cpio archive extraction.

The TEST-62-CPIO test should probably now be renamed TEST-62-SKIPCPIO... I'll squash that in with the next rebase.

Signed-off-by: David Disseldorp <[email protected]>

Individual test scripts may change working directory, so relative paths should be avoided. Signed-off-by: David Disseldorp <[email protected]>

Signed-off-by: David Disseldorp <[email protected]>

Crosvm's rust argument library is very small and simple, while still providing helpful functionality. It will be consumed by dracut-cpio in a subsequent commit. The unmodified, BSD licensed argument.rs source is lifted as-is from https://chromium.googlesource.com/chromiumos/platform/crosvm (release-R92-13982.B b6ae6517aeef9ae1e3a39c55b52f9ac6de8edb31). The one-line crosvm.rs wrapper is needed to ensure that crosvm::argument imports continue to work. Signed-off-by: David Disseldorp <[email protected]>

dracut-cpio is a minimal cpio archive creation utility written in Rust. It provides support for a minimal set of features needed to create performant and space-efficient initramfs archives: - "newc" archive format only - reproducible; inode numbers, uid/gid and mtime can be explicitly set - data segment copy-on-write reflinks + using Rust io::copy()'s native copy_file_range() support[1] + optional archive data segment alignment for optimal reflink use[2] - hardlink support - comprehensive tests asserting GNU cpio binary output compatibility 1. Rust io::copy() copy_file_range() rust-lang/rust#75272 2. Data segment alignment We're bending the newc spec a bit to inject zeros after the file path to provide data segment alignment. These zeros are accounted for in the namesize, but some applications may only expect a single zero-terminator (and 4 byte alignment). GNU cpio and Linux initramfs handle this fine as long as PATH_MAX isn't exceeded. Signed-off-by: David Disseldorp <[email protected]>

If configured with --enable-dracut-cpio, call cargo to build the dracut-cpio release binary. Signed-off-by: David Disseldorp <[email protected]>

The new dracut-cpio binary is capable of performing copy-on-write optimized initramfs archive creation, but due to the rust dependency isn't built / installed by default. This change adds a new "--enhanced-cpio" parameter for dracut which sees dracut-cpio called for archive creation instead of GNU cpio. Signed-off-by: David Disseldorp <[email protected]>

dracut-cpio already carries a bunch of unit tests covering compression and GNU cpio extraction. The purpose of these tests is to exercise the dracut.sh --enhanced-cpio code-paths as well as kernel cpio archive extraction. Signed-off-by: David Disseldorp <[email protected]>

ddiss · 2021-09-17T14:45:54Z

Changes since last version:

rebase atop f9c7dea
rename TEST-62-CPIO to TEST-62-SKIPCPIO, so that it's distinct from dracut-cpio tests

ddiss · 2021-09-20T12:30:33Z

Ping - is there anything I can do to move this forwards?

ddiss · 2021-09-29T13:59:51Z

I've updated the cover letter to include more recent benchmark results, which include XFS and Dracut zstd numbers for comparison.

ddiss · 2021-09-29T14:57:45Z

One other note is that we encountered what appears to be a bug in GRUB's Btrfs driver when reading initramfs images with a large number of shared extents: https://bugzilla.opensuse.org/show_bug.cgi?id=1190982 . A Btrfs developer is investigating the issue.
The GRUB issue was not seen with reflinked initramfs images on XFS.

ddiss · 2021-10-07T10:46:34Z

@haraldh ping - I'd really appreciate some feedback on whether this can be merged. Just to make it clear, there's no change to default behaviour here, the new functionality is only enabled if built with configure --enable-dracut-cpio and run with dracut --enhanced-cpio.

ddiss · 2021-10-15T13:50:51Z

@haraldh ping - I'd really appreciate some feedback on whether this can be merged. Just to make it clear, there's no change to default behaviour here, the new functionality is only enabled if built with configure --enable-dracut-cpio and run with dracut --enhanced-cpio.

Ping again... sorry to be a pain.

ddiss · 2021-10-18T11:02:57Z

One other note is that we encountered what appears to be a bug in GRUB's Btrfs driver when reading initramfs images with a large number of shared extents: https://bugzilla.opensuse.org/show_bug.cgi?id=1190982 . A Btrfs developer is investigating the issue.

FWIW, @adam900710 posted a patch which addresses the above GRUB Btrfs bug:
https://lists.gnu.org/archive/html/grub-devel/2021-10/msg00138.html

mwilck · 2021-10-20T18:54:10Z

FTR, I've done quite a few test runs with David's reflinking cpio, and have observed no issues, disk space savings on par with current compression. Here is an example:

root@bremer:boot> btrfs fi du i-xz-0 i-reflink i-zst-3
     Total   Exclusive  Set shared  Filename
  17.14MiB    17.14MiB       0.00B  i-xz-0
  40.00MiB    10.61MiB    29.39MiB  i-reflink
  18.94MiB    18.94MiB       0.00B  i-zst-3

The performance is also very good, it's faster than "zstd -3 -T0" in almost all cases.

johannbg

I dont see anything at this point why this cant be merged
@haraldh could you review this and add your feedback.

ddiss requested review from danimo, haraldh and johannbg as code owners June 8, 2021 23:55

github-actions bot added the test Issues related to testing label Jun 8, 2021

Firstyear reviewed Jun 9, 2021

View reviewed changes

rapid-cpio/Cargo.toml Outdated Show resolved Hide resolved

tpgxyz reviewed Jun 11, 2021

View reviewed changes

configure Show resolved Hide resolved

tpgxyz reviewed Jun 11, 2021

View reviewed changes

man/dracut.8.asc Show resolved Hide resolved

dracut.sh Outdated Show resolved Hide resolved

stale bot added the stale communication is stuck label Jul 12, 2021

johannbg removed the stale communication is stuck label Jul 17, 2021

ddiss force-pushed the dracut_cpio branch from 58eecd1 to 69843d2 Compare August 19, 2021 13:26

ddiss changed the title ~~improve initramfs image performance and efficiency via cpio reflinks~~ feat(cpio): improve initramfs image performance and efficiency via cpio reflinks Aug 27, 2021

ddiss force-pushed the dracut_cpio branch from 69843d2 to 92c9dc6 Compare August 27, 2021 16:34

ddiss force-pushed the dracut_cpio branch from 92c9dc6 to 97487d9 Compare September 2, 2021 16:02

ddiss force-pushed the dracut_cpio branch from fbe8d7f to 59d2e95 Compare September 3, 2021 19:24

ddiss added 5 commits September 17, 2021 16:22

ci(TEST-60-BONDBRIDGEVLANIFCFG): use toplevel Makefile

95a7754

Signed-off-by: David Disseldorp <[email protected]>

ci(test): export basedir and testdir as absolute paths

bff9aac

Individual test scripts may change working directory, so relative paths should be avoided. Signed-off-by: David Disseldorp <[email protected]>

ci(TEST-62-SKIPCPIO): add simple skipcpio test

2ee6eed

Signed-off-by: David Disseldorp <[email protected]>

ddiss added 3 commits September 17, 2021 16:26

feat(Makefile): cargo wrapper for dracut-cpio build

25cd05e

If configured with --enable-dracut-cpio, call cargo to build the dracut-cpio release binary. Signed-off-by: David Disseldorp <[email protected]>

ddiss force-pushed the dracut_cpio branch from 59d2e95 to d107e78 Compare September 17, 2021 14:43

ddiss mentioned this pull request Sep 17, 2021

Dracut cpio-reflink feature detection ddiss/dracut#1

Closed

johannbg approved these changes Nov 7, 2021

View reviewed changes

haraldh approved these changes Nov 24, 2021

View reviewed changes

haraldh merged commit 8104bf0 into dracutdevs:master Nov 24, 2021

This was referenced Dec 6, 2021

[RFC] feature: use reflinks for extent sharing between initramfs source and archive data #1141

Closed

feature: copy or reflink cpio data directly from source instead of staging in tmpdir #1662

Open

LaszloGombos mentioned this pull request Apr 5, 2024

Enable test for dracut-cpio dracut-ng/dracut-ng#106

Closed

ddiss mentioned this pull request Apr 11, 2024

test(DRACUT-CPIO): enable building dracut-cpio when cargo is installed dracut-ng/dracut-ng#107

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(cpio): improve initramfs image performance and efficiency via cpio reflinks #1531

feat(cpio): improve initramfs image performance and efficiency via cpio reflinks #1531

ddiss commented Jun 8, 2021 •

edited

Loading

johannbg commented Jun 12, 2021

stale bot commented Jul 12, 2021

johannbg commented Jul 17, 2021

ddiss commented Jul 26, 2021

the8472 commented Jul 29, 2021

ddiss commented Jul 29, 2021

ddiss commented Aug 19, 2021

ddiss commented Aug 27, 2021

TomasTomecek commented Aug 28, 2021

ddiss commented Sep 2, 2021

ddiss commented Sep 3, 2021

ddiss commented Sep 17, 2021

ddiss commented Sep 20, 2021

ddiss commented Sep 29, 2021

ddiss commented Sep 29, 2021

ddiss commented Oct 7, 2021

ddiss commented Oct 15, 2021

ddiss commented Oct 18, 2021

mwilck commented Oct 20, 2021

johannbg left a comment

feat(cpio): improve initramfs image performance and efficiency via cpio reflinks #1531

feat(cpio): improve initramfs image performance and efficiency via cpio reflinks #1531

Conversation

ddiss commented Jun 8, 2021 • edited Loading

Changes

Performance

Checklist

johannbg commented Jun 12, 2021

stale bot commented Jul 12, 2021

johannbg commented Jul 17, 2021

ddiss commented Jul 26, 2021

the8472 commented Jul 29, 2021

ddiss commented Jul 29, 2021

ddiss commented Aug 19, 2021

ddiss commented Aug 27, 2021

TomasTomecek commented Aug 28, 2021

ddiss commented Sep 2, 2021

ddiss commented Sep 3, 2021

ddiss commented Sep 17, 2021

ddiss commented Sep 20, 2021

ddiss commented Sep 29, 2021

ddiss commented Sep 29, 2021

ddiss commented Oct 7, 2021

ddiss commented Oct 15, 2021

ddiss commented Oct 18, 2021

mwilck commented Oct 20, 2021

johannbg left a comment

Choose a reason for hiding this comment

ddiss commented Jun 8, 2021 •

edited

Loading