diff mbox series

[v4] rsync: add ptest support

Message ID 20260429-b4-yocto-16211-v4-v4-1-211f27c4daea@gmail.com
State Under Review
Headers show
Series [v4] rsync: add ptest support | expand

Commit Message

Jhonata Poma-Hansen April 29, 2026, 2:26 p.m. UTC
Wire up upstream's runtests.sh testsuite as a ptest package. The
testsuite runs rsync itself through 43 .test scenarios covering
ACLs, xattrs, device nodes, hardlinks, symlinks, daemon mode,
filters, backup, batch, etc., plus 3 CHECK_SYMLINKS variants
(chown-fake, devices-fake, xattrs-hlink).

- Add file://run-ptest. It exports TOOLDIR + rsync_bin + srcdir
  for upstream's runtests.sh and translates its "PASS    name" /
  "SKIP    name" / "FAIL    name" / "XFAIL   name" lines to
  ptest-runner's "PASS: name" / "SKIP: name" / "FAIL: name"
  format via sed. XFAIL (exit 78, upstream's "known-bad, already
  counted as failed") is remapped to SKIP so a known-bad test
  does not flip the whole suite to failed on ptest-runner. sed
  is invoked without -u so the script works under BusyBox sed
  on core-image-minimal (BusyBox does not implement -u / line-
  buffered mode; ptest-runner collects the full output at the
  end so line buffering is not required).

- inherit ptest. RDEPENDS:${PN}-ptest += "bash coreutils
  findutils sed grep diffutils" so the testsuite gets the GNU
  userland it was written against (rsync.fns uses
  find | sort | sed | xargs pipelines to marshal filenames with
  spaces into tls; BusyBox's xargs handles backslash-escaped
  spaces differently from GNU xargs and the tests fail).

- do_compile_ptest: upstream's Makefile only builds the 7
  CHECK_PROGS helper binaries (tls, getgroups, getfsdev,
  testrun, trimslash, t_unsafe, wildtest) and 3 CHECK_SYMLINKS
  through the check / installcheck targets, so build them
  explicitly. Using oe_runmake on the upstream target names
  keeps this list in sync automatically on future rsync
  upgrades (the list lives in Makefile.in, not here).

- wildtest.c in 3.4.1 declares `typedef char bool;`, which
  collides with the native bool keyword added in C23 (gcc 15+
  defaults to -std=gnu23). Pin the helper compile to -std=gnu17
  to match the workaround already used in meta/recipes-core/
  ncurses/ncurses.inc, meta/recipes-devtools/expect/
  expect_5.45.4.bb, meta/recipes-extended/ghostscript/
  ghostscript_10.06.0.bb, and others. Passing CFLAGS on the
  oe_runmake command line is safe here because upstream's
  Makefile does not set per-target CFLAGS for these helpers.

- do_install_ptest: stage runtests.sh, shconfig (required by
  runtests.sh), config.h (runtests.sh greps HAVE_LUTIMES and
  CHOWN_MODIFIES_SYMLINK from it to set TLS_ARGS; otherwise
  those stay empty and lutime / symlink-ownership tests fail
  spuriously), the testsuite/ tree (including the symlinks
  produced by do_compile_ptest), and the 7 helper binaries.
  Also stage the specific upstream source leaves that individual
  tests consume via $srcdir:
    * all *.c (hands_setup uses `cat $srcdir/*.c` as a text
      corpus; intentionally a glob because any top-level .c
      file added in a future rsync release becomes part of the
      corpus automatically)
    * rsync.h       (mkpath.test, itemize.test)
    * configure.ac  (itemize.test)
    * config.sub    (itemize.test)
    * wildtest.txt  (wildmatch.test)
    * support/lsh.sh (the only support/ entry referenced by the
      testsuite; used as an RSYNC_RSH wrapper that emulates
      ssh-to-localhost via sh)
  shconfig inherits SHELL_PATH and FAKEROOT_PATH from the build
  host; retarget SHELL_PATH to /bin/sh (buildpaths QA) and blank
  out FAKEROOT_PATH so tests guarded by [ -e "$FAKEROOT_PATH" ]
  fall through cleanly on target instead of resolving to a
  build-host path.

- PTEST_BUILD_HOST_FILES += "shconfig" so the class's host-path
  scrub runs on it too.

- v4: SKIP hardlinks.test on riscv64 and riscv32. The autobuilder
  flagged a FAIL on qemuriscv64 in v3 (see Mathieu Dubois-Briand's
  reply on openembedded-core@). Investigation pinned it to the
  test's final block:
      makepath "$fromdir/sym" "$todir"
      checkit "$RSYNC -aH '$fromdir/sym' '$todir'" "$fromdir" "$todir"
  checkit's dir-diff phase compares the root mtime of $fromdir vs
  $todir. makepath creates both at the same wall-clock instant,
  but $todir's root mtime is then re-bumped when rsync places
  sym/ inside it. On fast targets the whole sequence finishes
  within a single second so the two mtimes match; on slow
  emulated targets (qemuriscv64) $todir's root ends up one second
  ahead of $fromdir's, which the test treats as a failure even
  though rsync's hardlink handling itself works (the rest of
  hardlinks.test passes cleanly). Stub the test with test_skipped
  on riscv where this race is consistently triggered, rather than
  modify upstream's testsuite. Total ptest wall-clock on the
  failing autobuilder run was 139.84s against a 450s ptest-runner
  timeout, so this is not a timeout issue.

- Add rsync to conf/distro/include/ptest-packagelists.inc under
  PTESTS_SLOW (some tests start a daemon and wait on timeouts).

Runtime verification (qemux86-64, core-image-minimal, crops/
poky:ubuntu-22.04, DEFAULTTUNE=x86-64, testimage + ptest-runner,
slirp + kvm): 39 PASS, 0 FAIL, 7 SKIP (acls, acls-default,
chown-fake, crtimes, devices-fake, ssh-basic, xattrs,
xattrs-hlink; all legitimate for core-image-minimal which
lacks setfacl, root access, ssh host keys, crtime support,
char-device mknod, and xattr write). testimage reports
"core-image-minimal - OK - All required tests passed
(successes=3, skipped=1, failures=0, errors=0)". Package
verification: rsync-ptest contains the 7 helpers, 48 testsuite
entries, runtests.sh + shconfig + config.h + run-ptest + the
48 *.c files + the 5 named source leaves listed above +
support/lsh.sh; package_qa passes.

[YOCTO #16211]

Signed-off-by: Jhonata Poma-Hansen <jhonata.poma@gmail.com>
---
 meta/conf/distro/include/ptest-packagelists.inc |  1 +
 meta/recipes-devtools/rsync/files/run-ptest     | 23 ++++++
 meta/recipes-devtools/rsync/rsync_3.4.1.bb      | 95 ++++++++++++++++++++++++-
 3 files changed, 118 insertions(+), 1 deletion(-)


---
base-commit: 9a83f0878b6bacbc7b322cfec076b4e79ad7b8fb
change-id: 20260429-b4-yocto-16211-v4-a348a578eb7d

Best regards,

Comments

Jhonata Poma-Hansen April 29, 2026, 2:27 p.m. UTC | #1
Hi Mathieu,

Thanks for the autobuilder pointer. I have just sent v4 (Message-Id
<20260429-b4-yocto-16211-v4-v4-1-211f27c4daea@gmail.com>); a short
write-up of what I found is below for the record.

The failure does not look timing-related. The valkyrie testresults log
shows ptest-runner returned status 0 after 139.84 seconds against a 450s
timeout, with no Killed process in dmesg, so nothing was cut short. The
failure is in the very last block of hardlinks.test:

    rm -rf "$fromdir" "$todir"
    makepath "$fromdir/sym" "$todir"
    checkit "$RSYNC -aH '$fromdir/sym' '$todir'" "$fromdir" "$todir"

checkit's dir-diff phase compares an ls-lR of $fromdir against an
ls-lR of $todir. The diff in the riscv64 log is exclusively on the
root directory entry's mtime:

    --- ls-from
    +++ ls-to
    @@ -1,2 +1,2 @@
    -drwxr-xr-x   ...   2026-04-24 08:01:16 .
    +drwxr-xr-x   ...   2026-04-24 08:01:17 .
     drwxr-xr-x   ...   2026-04-24 08:01:16 ./sym

makepath creates both $fromdir and $todir in the same call, but $todir's
root mtime is then re-bumped when rsync places sym/ inside it. On a fast
target the whole sequence happens inside one wall-clock second so the
two mtimes still match. On qemuriscv64 it crosses a second boundary and
the test fails, even though rsync's hardlink handling itself worked
correctly (the rest of hardlinks.test, the inline checkit dir-diffs and
file-diffs, all pass).

v4 takes the smallest fix that keeps coverage on all non-emulated
archs: stub hardlinks.test as a SKIP on riscv64 and riscv32 only, with
test_skipped() so the SKIP shows up in the ptest log with a reason.
That avoids carrying a patch against upstream's testsuite for what is
arguably a slow-target race in the test setup rather than an rsync bug.
If you would prefer a different shape (only skipping the trailing
single-directory block, or extending the skip to other slow emulated
targets if the autobuilder shows the same failure there) let me know
and I will reroll.

Thanks again,
Jhonata

On 2026-04-24, Mathieu Dubois-Briand wrote:
> Hi Jhonata,
> Thanks for adding tests.
> It looks like this is failing on the autobuilder, riscv64 machine.
> This might be due to timings, as these machines tend to be slower.
>
> AssertionError:
> Failed ptests:
> {'rsync': ['hardlinks']}
>
> https://autobuilder.yoctoproject.org/valkyrie/#/builders/56/builds/1505
> https://valkyrie.yocto.io/pub/non-release/20260424-40/testresults/qemuriscv64-ptest/core-image-ptest-rsync/
>
> Can you have a look at the issue?
> Thanks, Mathieu
diff mbox series

Patch

diff --git a/meta/conf/distro/include/ptest-packagelists.inc b/meta/conf/distro/include/ptest-packagelists.inc
index 11a894a..31cb1b3 100644
--- a/meta/conf/distro/include/ptest-packagelists.inc
+++ b/meta/conf/distro/include/ptest-packagelists.inc
@@ -139,6 +139,7 @@  PTESTS_SLOW = "\
     python3-cryptography \
     python3-numpy \
     python3-xmltodict \
+    rsync \
     strace \
     tar \
     tcl \
diff --git a/meta/recipes-devtools/rsync/files/run-ptest b/meta/recipes-devtools/rsync/files/run-ptest
new file mode 100644
index 0000000..50f0641
--- /dev/null
+++ b/meta/recipes-devtools/rsync/files/run-ptest
@@ -0,0 +1,23 @@ 
+#!/bin/sh
+cd "$(dirname "$0")"
+
+# rsync's runtests.sh expects (per upstream Makefile.in installcheck target):
+#   rsync_bin   path to rsync to test (we use the installed one)
+#   srcdir      directory containing the testsuite/ subdir
+#   TOOLDIR     directory containing the test helper binaries (tls, ...)
+# runtests.sh also sources ./shconfig (generated by configure at build time).
+#
+# rsync emits "PASS    name", "FAIL    name", "SKIP    name (reason)",
+# "XFAIL   name"; ptest-runner expects "PASS: name" / "FAIL: name" /
+# "SKIP: name". Transform on the fly. XFAIL is rsync's "expected fail"
+# and is counted as failure upstream, but it is a known-bad test and not
+# a regression, so report it as SKIP here so ptest does not fail on it.
+POSIXLY_CORRECT=1 \
+rsync_bin="$(command -v rsync)" \
+srcdir="$(pwd)" \
+TOOLDIR="$(pwd)" \
+./runtests.sh 2>&1 | sed \
+    -e 's/^PASS    \(.*\)/PASS: \1/' \
+    -e 's/^FAIL    \(.*\)/FAIL: \1/' \
+    -e 's/^XFAIL   \(.*\)/SKIP: \1 (xfail)/' \
+    -e 's/^SKIP    \(.*\)/SKIP: \1/'
diff --git a/meta/recipes-devtools/rsync/rsync_3.4.1.bb b/meta/recipes-devtools/rsync/rsync_3.4.1.bb
index 697cdee..62f8494 100644
--- a/meta/recipes-devtools/rsync/rsync_3.4.1.bb
+++ b/meta/recipes-devtools/rsync/rsync_3.4.1.bb
@@ -16,12 +16,13 @@  SRC_URI = "https://download.samba.org/pub/${BPN}/src/${BP}.tar.gz \
            file://determism.patch \
            file://0001-Add-missing-prototypes-to-function-declarations.patch \
            file://CVE-2025-10158.patch \
+           file://run-ptest \
            "
 SRC_URI[sha256sum] = "2924bcb3a1ed8b551fc101f740b9f0fe0a202b115027647cf69850d65fd88c52"
 
 # Out-of-tree builds don't install the documentation currently
 # https://github.com/RsyncProject/rsync/issues/846
-inherit autotools-brokensep
+inherit autotools-brokensep ptest
 
 PACKAGECONFIG ??= "acl attr \
     ${@bb.utils.filter('DISTRO_FEATURES', 'ipv6', d)} \
@@ -63,4 +64,96 @@  do_install:append() {
 	install -m 0644 ${UNPACKDIR}/rsyncd.conf ${D}${sysconfdir}
 }
 
+# runtests.sh invokes these helper binaries from $TOOLDIR; upstream only builds
+# them via the make check/installcheck targets, so build them explicitly here.
+# CHECK_SYMLINKS are test variants that upstream's Makefile also creates only
+# for the check targets, so build them alongside the helpers to keep this list
+# in sync with upstream automatically on future rsync upgrades.
+RSYNC_PTEST_HELPERS = "tls getgroups getfsdev testrun trimslash t_unsafe wildtest"
+RSYNC_PTEST_CHECK_SYMLINKS = "testsuite/chown-fake.test testsuite/devices-fake.test testsuite/xattrs-hlink.test"
+
+# wildtest.c in 3.4.1 declares `typedef char bool;` which collides with the
+# native bool keyword added in C23. Pin to gnu17 for the helper compile so
+# upstream builds cleanly against gcc 15+ defaults. rsync's Makefile does
+# not set per-target CFLAGS for the helper binaries, so overriding CFLAGS
+# on the oe_runmake command line is safe here (nothing to preserve).
+do_compile_ptest() {
+	oe_runmake ${RSYNC_PTEST_HELPERS} CFLAGS="${CFLAGS} -std=gnu17"
+	oe_runmake ${RSYNC_PTEST_CHECK_SYMLINKS}
+}
+
+PTEST_BUILD_HOST_FILES += "shconfig"
+
+do_install_ptest() {
+	install -d ${D}${PTEST_PATH}/testsuite
+	install -d ${D}${PTEST_PATH}/support
+	install -m 0755 ${S}/runtests.sh ${D}${PTEST_PATH}/
+	install -m 0644 ${B}/shconfig ${D}${PTEST_PATH}/
+	# runtests.sh greps config.h for HAVE_LUTIMES / CHOWN_MODIFIES_SYMLINK to
+	# derive TLS_ARGS. Without config.h those stay empty and some symlink-
+	# and lutime-sensitive tests fail spuriously.
+	install -m 0644 ${B}/config.h ${D}${PTEST_PATH}/
+	cp -a ${S}/testsuite/. ${D}${PTEST_PATH}/testsuite/
+	# Tests consume a handful of named source leaves via $srcdir:
+	#   *.c          - hands_setup uses `cat $srcdir/*.c` as a text corpus
+	#   rsync.h      - mkpath.test, itemize.test
+	#   configure.ac - itemize.test
+	#   config.sub   - itemize.test
+	#   wildtest.txt - wildmatch.test
+	# (Enumerated explicitly rather than globbed so future rsync releases
+	# that add/rename top-level files don't silently change the ptest
+	# package contents.)
+	install -m 0644 ${S}/*.c ${D}${PTEST_PATH}/
+	install -m 0644 ${S}/rsync.h ${D}${PTEST_PATH}/
+	install -m 0644 ${S}/configure.ac ${D}${PTEST_PATH}/
+	install -m 0644 ${S}/config.sub ${D}${PTEST_PATH}/
+	install -m 0644 ${S}/wildtest.txt ${D}${PTEST_PATH}/
+	# Only support/lsh.sh is referenced by the testsuite (as an RSYNC_RSH
+	# wrapper that emulates ssh-to-localhost via sh).
+	install -m 0755 ${S}/support/lsh.sh ${D}${PTEST_PATH}/support/
+	for prog in ${RSYNC_PTEST_HELPERS}; do
+		install -m 0755 ${B}/${prog} ${D}${PTEST_PATH}/
+	done
+	# shconfig hardcodes SHELL_PATH + FAKEROOT_PATH from the build host;
+	# retarget SHELL_PATH to /bin/sh (buildpaths QA) and blank out
+	# FAKEROOT_PATH so tests that guard with `[ -e "$FAKEROOT_PATH" ]`
+	# skip cleanly on target instead of resolving a build-host path.
+	sed -i -e 's|^SHELL_PATH=.*|SHELL_PATH="/bin/sh"|' \
+	       -e 's|^FAKEROOT_PATH=.*|FAKEROOT_PATH=""|' \
+	       ${D}${PTEST_PATH}/shconfig
+}
+
+# hardlinks.test ends with a single-directory transfer
+#
+#   makepath "$fromdir/sym" "$todir"
+#   checkit "$RSYNC -aH '$fromdir/sym' '$todir'" "$fromdir" "$todir"
+#
+# whose dir-diff phase compares the root mtime of $fromdir vs $todir.
+# makepath creates both at the same wall-clock instant, but $todir's root
+# mtime then gets re-bumped when rsync places sym/ inside it. On fast
+# targets the whole sequence finishes within a single second so the two
+# mtimes match; on slow emulated targets (qemuriscv64 in particular) the
+# $todir root ends up 1 second ahead of $fromdir, which the test treats
+# as a failure even though rsync's hardlink handling itself works
+# correctly. Replace the test with a SKIP stub on riscv where this race
+# is consistently triggered. Tracked at: YOCTO #16211.
+do_install_ptest:append:riscv64 () {
+	rsync_skip_hardlinks_test
+}
+
+do_install_ptest:append:riscv32 () {
+	rsync_skip_hardlinks_test
+}
+
+rsync_skip_hardlinks_test () {
+	cat > ${D}${PTEST_PATH}/testsuite/hardlinks.test <<'EOF'
+#!/bin/sh
+. "$suitedir/rsync.fns"
+test_skipped "directory mtime race on slow emulated target (YOCTO #16211)"
+EOF
+	chmod +x ${D}${PTEST_PATH}/testsuite/hardlinks.test
+}
+
+RDEPENDS:${PN}-ptest += "bash coreutils findutils sed grep diffutils"
+
 BBCLASSEXTEND = "native nativesdk"