rsync 3.5.1: read the protocol 33 block stat, and stop truncating byte totals

[ralyodio]

What

rsync 3.5.1 shipped 2026-09-21 with protocol 33, which adds one line to --stats:

Number of 4 KiB logical blocks touched: 857

printed between Literal data and Matched data, and only when both ends negotiate 33.

DiskPush already ran fine on 3.5.1 — every capability check is a lower bound, and the parser passes lines it does not recognise through to the log — but that number was being dropped on the floor. This reads it, and covers both the version distros ship today (3.4.x) and the latest.

• RsyncStats.logicalBlocksTouched, parsed from the real label, verbatim from rsync's main.c.
• A blockStats capability taken from the negotiated protocol rather than the version string, falling back to the version only for a banner that prints no protocol line.
• Intersected like every other capability, so a mixed 3.5/3.4 pair reports null rather than a number only one end can produce.
• The CLI prints Blocks touched: 856 x 4 KiB when it is available and stays silent when it is not. stats is now in the --json payload.

The bug this turned up

Verifying against a real 3.5.1 build exposed something unrelated to 3.5.1 and older than it. rsync prints byte totals with human_num, so under --human-readable — which DiskPush passes by default — the output is Total bytes sent: 3.50M. The stats patterns only read [\d,]+, so that parsed as 3.

Every byte field was affected on every rsync version:

field	was	now
totalBytesSent	3	3500000
literalBytes	3	3500000
speedup	null	1

Byte fields now read the unit suffix back through the same conversion the progress line already used. Counts keep their exact comma_num parsing, so the new block stat is unaffected either way.

I fixed this rather than only reporting it because this PR puts stats into the --json output, and publishing knowingly wrong numbers in a new contract seemed worse than the extra diff. Happy to split it out if you would rather.

Verification

Not just unit tests — built rsync 3.5.1 from source and ran both binaries for real.

3.5.1: protocol 33 | blockStats true  | mkpath true | secludedByDefault true
3.4.1: protocol 32 | blockStats false | mkpath true | secludedByDefault true
mixed ->  protocol 32 | blockStats false

Parsed from genuine --stats output:

3.5.1: {"filesTotal":5,"filesTransferred":3,"totalBytesSent":3500000,"literalBytes":3500000,"logicalBlocksTouched":857,"speedup":1}
3.4.1: {"filesTotal":5,"filesTransferred":3,"totalBytesSent":3500000,"literalBytes":3500000,"logicalBlocksTouched":null,"speedup":1}

• 723 tests pass on stock 3.4.1.
• The full live suite (12 tests) passes against real 3.5.1, including the resume test — which exercises exactly the partial-directory validation 3.5.1 tightened. --partial-dir resume is unaffected.
• CLI end to end on both: the line appears on 3.5.1, is absent on 3.4.1.
• The new live test is version-aware, so it asserts the right branch on whatever rsync the machine has.

Note for a follow-up

3.5.1 also changed --files-from paths to be treated as operator-supplied rather than transfer-root-relative. DiskPush passes --files-from straight through, so the semantic shift is upstream's, but it could change path resolution for anyone using that flag on 3.5.1. Not addressed here.

🤖 Generated with Claude Code