Let CI give up on a dnf mirror that is slow or hangs

.github/workflows/test.yml

@@ -60,13 +60,14 @@ jobs:
             test ! -e /etc/apt/sources.list.d/azentrius.sources'
           for image in fedora:43 rockylinux/rockylinux:9 rockylinux/rockylinux:10 almalinux:9 almalinux:10; do
             echo "::group::$image"
-            docker run --rm -v "$AZPKG_ROOT:/repo:ro" -v "$PWD/repository/azentrius.asc:/azentrius.asc:ro" "$image" sh -ec '
+            docker run --rm -v "$AZPKG_ROOT:/repo:ro" -v "$PWD/repository/azentrius.asc:/azentrius.asc:ro" \
+              -v "$PWD/tools/dnf.sh:/dnf.sh:ro" "$image" sh -ec '
               . /etc/os-release
               echo "$PRETTY_NAME, $(rpm --version)"
               rpm --import /azentrius.asc
               rpm --import /repo/test.asc
               printf "[test]\nname=test\nbaseurl=file:///repo/rpm\ngpgcheck=1\n" > /etc/yum.repos.d/test.repo
-              dnf install -y -q azentrius-repository
+              sh /dnf.sh install -y -q azentrius-repository
               rpm -K /repo/rpm/azentrius-repository-*.rpm | grep -q "digests signatures OK"
               test -f /usr/share/swcatalog/xml/azentrius.xml
               grep -q "baseurl=https://packages.azentrius.com/rpm" /etc/yum.repos.d/azentrius.repo

README.md

@@ -118,7 +118,9 @@ server tool and installs `azentrius-repository` from it on Debian 13, Fedora 43,
 10 and AlmaLinux 9 and 10; the RPM systems import the repository's real key first.
 `tools/build-packages.sh` is the same file in every Azentrius add-on. `tools/apt-update.sh` runs
 `apt-get update` on the CI runner with time limits and three tries, so that a mirror that hangs
-costs minutes instead of the job; it is the same file as in Cron Jobs.
+costs minutes instead of the job; it is the same file as in Cron Jobs. `tools/dnf.sh`, also from
+there, runs dnf in the test containers with time limits and three tries, the repository data
+first, and lets dnf leave a mirror slower than 100 kB/s.
 
 Releases: raise `VERSION`, add a section to `CHANGELOG.md`, merge, then Actions › release › Run
 workflow on `main`.

tools/dnf.sh

@@ -0,0 +1,39 @@
+#!/bin/sh
+# dnf in a CI container, as root, that leaves a slow mirror and ends a try that hangs, instead of
+# waiting until the job times out: on AlmaLinux the first dnf install of a job, which downloads
+# the repository data, once took almost 14 minutes without a word under -q (normally well under a
+# minute), and once ran into the job's 20-minute limit.
+#
+#   sh /dnf.sh install -y -q PACKAGE...   the repository data first, then the command
+#   sh /dnf.sh makecache                  only the repository data, for a dnf of its own after it
+#
+# dnf leaves a mirror that sends less than 100 kB/s for 20 seconds and takes the next one (its own
+# defaults: 1 kB/s, 30 seconds); the two settings go into /etc/dnf/dnf.conf and stay for every later
+# dnf in the container. The repository data comes first and without -q, so the log shows each
+# repository's size and speed. Each try may take 180 seconds; a failed or hung one gets a second
+# and a third, 15 seconds apart.
+# SPDX-License-Identifier: Apache-2.0
+set -u
+[ $# -gt 0 ] || { echo "usage: sh dnf.sh makecache | install -y -q PACKAGE..." >&2; exit 64; }
+conf=/etc/dnf/dnf.conf
+grep -qx 'minrate=100000' "$conf" 2> /dev/null ||
+    sed -i -e '/^\[main\]/a minrate=100000' -e '/^\[main\]/a timeout=20' "$conf"
+grep -qx 'minrate=100000' "$conf" || { echo "::error::no [main] section in $conf"; exit 1; }
+
+tries() {
+    for try in 1 2 3; do
+        rc=0
+        timeout --kill-after=10 180 dnf "$@" || rc=$?
+        [ "$rc" != 0 ] || return 0
+        case $rc in
+            124 | 137) why="no end within 180 seconds" ;;
+            *) why="exit code $rc" ;;
+        esac
+        echo "::warning::dnf $1 failed ($why, try $try of 3)"
+        [ "$try" -lt 3 ] || { echo "::error::dnf $1 failed three times"; exit "$rc"; }
+        sleep 15
+    done
+}
+
+tries makecache
+[ "$1" = makecache ] || tries "$@"