On Thu, Aug 14, 2025 at 03:18:05PM +1000, David Disseldorp wrote: > As described in buffer-format.rst, the existing initramfs.c extraction > logic works fine if the cpio filename field is padded out with trailing > zeros, with a caveat that the padded namesize can't exceed PATH_MAX. > > Add filename zero-padding logic to gen_init_cpio, which can be triggered > via the new -a <data_align> parameter. Performance and storage > utilization is improved for Btrfs and XFS workloads, as copy_file_range > can reflink the entire source file into a filesystem block-size aligned > destination offset within the cpio archive. > > Btrfs benchmarks run on 6.15.8-1-default (Tumbleweed) x86_64 host: > > truncate --size=2G /tmp/backing.img > > /sbin/mkfs.btrfs /tmp/backing.img > ... > Sector size: 4096 (CPU page size: 4096) > ... > > sudo mount /tmp/backing.img mnt > > sudo chown $USER mnt > > cd mnt > mnt> dd if=/dev/urandom of=foo bs=1M count=20 && cat foo >/dev/null > ... > mnt> echo "file /foo foo 0755 0 0" > list > mnt> perf stat -r 10 gen_init_cpio -o unaligned_btrfs list > ... > 0.023496 +- 0.000472 seconds time elapsed ( +- 2.01% ) > > mnt> perf stat -r 10 gen_init_cpio -o aligned_btrfs -a 4096 list > ... > 0.0010010 +- 0.0000565 seconds time elapsed ( +- 5.65% ) > > mnt> /sbin/xfs_io -c "fiemap -v" unaligned_btrfs > unaligned_btrfs: > EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS > 0: [0..40967]: 695040..736007 40968 0x1 > mnt> /sbin/xfs_io -c "fiemap -v" aligned_btrfs > aligned_btrfs: > EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS > 0: [0..7]: 26768..26775 8 0x0 > 1: [8..40967]: 269056..310015 40960 0x2000 > 2: [40968..40975]: 26776..26783 8 0x1 > mnt> /sbin/btrfs fi du unaligned_btrfs aligned_btrfs > Total Exclusive Set shared Filename > 20.00MiB 20.00MiB 0.00B unaligned_btrfs > 20.01MiB 8.00KiB 20.00MiB aligned_btrfs > > XFS benchmarks run on same host: > > sudo umount mnt && rm /tmp/backing.img > > truncate --size=2G /tmp/backing.img > > /sbin/mkfs.xfs /tmp/backing.img > ... > = reflink=1 ... > data = bsize=4096 blocks=524288, imaxpct=25 > ... > > sudo mount /tmp/backing.img mnt > > sudo chown $USER mnt > > cd mnt > mnt> dd if=/dev/urandom of=foo bs=1M count=20 && cat foo >/dev/null > ... > mnt> echo "file /foo foo 0755 0 0" > list > mnt> perf stat -r 10 gen_init_cpio -o unaligned_xfs list > ... > 0.011069 +- 0.000469 seconds time elapsed ( +- 4.24% ) > > mnt> perf stat -r 10 gen_init_cpio -o aligned_xfs -a 4096 list > ... > 0.001273 +- 0.000288 seconds time elapsed ( +- 22.60% ) > > mnt> /sbin/xfs_io -c "fiemap -v" unaligned_xfs > unaligned_xfs: > EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS > 0: [0..40967]: 106176..147143 40968 0x0 > 1: [40968..65023]: 147144..171199 24056 0x801 > mnt> /sbin/xfs_io -c "fiemap -v" aligned_xfs > aligned_xfs: > EXT: FILE-OFFSET BLOCK-RANGE TOTAL FLAGS > 0: [0..7]: 120..127 8 0x0 > 1: [8..40967]: 192..41151 40960 0x2000 > 2: [40968..40975]: 236728..236735 8 0x0 > 3: [40976..106495]: 236736..302255 65520 0x801 > > The alignment is best-effort; a stderr message is printed if alignment > can't be achieved due to PATH_MAX overrun, with fallback to non-padded > filename. This allows it to still be useful for opportunistic alignment, > e.g. on aarch64 Btrfs with 64K block-size. Alignment failure messages > provide an indicator that reordering of the cpio-manifest may be > beneficial. > > Archive read performance for reflinked initramfs images may suffer due > to the effects of fragmentation, particularly on spinning disks. To > mitigate excessive fragmentation, files with lengths less than > data_align aren't padded. > > Signed-off-by: David Disseldorp <ddiss@xxxxxxx> > --- > usr/gen_init_cpio.c | 50 ++++++++++++++++++++++++++++++++++----------- > 1 file changed, 38 insertions(+), 12 deletions(-) Thanks! Testing with a massively oversized initramfs (600MB) was fun: from 2:44 down to 38s. Questions that pop up in my mind: Now, how can we make other benefit from this? Might it make sense to introduce a kconfig variable for initramfs alignment -- even though this is just a build-time optimisation of few seconds? Kind regards, Nicolas
