fix(bootstrap): preflight 探测下载源重试三次并以 TLS 握手为准，一次网络抖动不再拒绝安装

deploy/bootstrap.sh

@@ -1314,18 +1314,27 @@ preflight_hosts() {
   fi
 }
 
-# host_reachable reports whether an HTTPS connection to $1 can be made: any answer
-# counts, a 404 included. Without curl (a minimal image, before install_base) a bare
-# TCP connect stands in, unless a proxy is configured, which only curl would use.
+# host_reachable reports whether an HTTPS connection to $1 can be made: a TLS handshake that
+# completes, which any answer (a 404 included) comes after, and which a server that then sits
+# on the request has made too. It tries three times, two seconds apart, as the downloads it
+# stands for retry: one dropped probe must not stop an install. Without curl (a minimal
+# image, before install_base) a bare TCP connect stands in, unless a proxy is configured,
+# which only curl would use.
 host_reachable() {
-  if command -v curl >/dev/null 2>&1; then
-    local code
-    code="$(curl -s -o /dev/null --connect-timeout 5 --max-time 15 -w '%{http_code}' "https://$1/" 2>/dev/null)" || true
-    [ -n "$code" ] && [ "$code" != 000 ]
-    return
-  fi
-  [ -z "${https_proxy:-}${HTTPS_PROXY:-}" ] || return 0
-  timeout 5 bash -c 'exec 3<>"/dev/tcp/$0/443"' "$1" 2>/dev/null
+  local try tls
+  for try in 1 2 3; do
+    [ "$try" = 1 ] || sleep 2
+    if command -v curl >/dev/null 2>&1; then
+      # The handshake's time, 0.000000 until one completes, whatever curl then exits with.
+      tls="$(curl -s -o /dev/null --connect-timeout 5 --max-time 15 -w '%{time_appconnect}' "https://$1/" 2>/dev/null)" || true
+      [ -n "${tls//[0.]/}" ] && return 0
+    elif [ -n "${https_proxy:-}${HTTPS_PROXY:-}" ]; then
+      return 0
+    elif timeout 5 bash -c 'exec 3<>"/dev/tcp/$0/443"' "$1" 2>/dev/null; then
+      return 0
+    fi
+  done
+  return 1
 }
 
 preflight_outbound() {

deploy/bootstrap_test.sh

@@ -2746,11 +2746,21 @@ run_pf() {
         *) printf "%s\n" "${PF_ADDRS:-}" ;;
       esac
     }
+    # curl -w "%{time_appconnect}": a host in PF_DOWN never answers, one in PF_FLAKY drops
+    # its first probe, one in PF_STALL completes TLS and then times out. Each probe of those
+    # is counted in $PF_ROOT/tries.<host>.
     curl() {
       local host="${!#}"
       host="${host#https://}"; host="${host%/}"
-      case " ${PF_DOWN:-} " in *" ${host} "*) echo 000 ;; *) echo 404 ;; esac
+      case " ${PF_DOWN:-} ${PF_FLAKY:-} ${PF_STALL:-} " in *" ${host} "*) printf x >> "$PF_ROOT/tries.${host}" ;; esac
+      case " ${PF_FLAKY:-} " in *" ${host} "*)
+        [ "$(cat "$PF_ROOT/tries.${host}")" = x ] && printf 0.000000 && return 7
+        printf 0.300000; return ;;
+      esac
+      case " ${PF_STALL:-} " in *" ${host} "*) printf 0.481676; return 28 ;; esac
+      case " ${PF_DOWN:-} " in *" ${host} "*) printf 0.000000; return 7 ;; *) printf 0.300000 ;; esac
     }
+    sleep() { echo "SLEEP: $*"; }
     bootstrap_from_tui() { [ -n "${PF_TUI:-}" ]; }
     '"$(awk '/^use_release_binary\(\) \{/,/^}/' "$BS")"'
     FELIS_GAME_PORT=25565 FELIS_PANEL_NODEPORT=30443 REGISTRY_URL=registry.felis.svc:5000 PG_HOST_PORT=15432
@@ -2834,6 +2844,21 @@ expect "and the install goes on" "WENT ON" "$out"
 
 out="$(PF_DOWN="github.com fill-data.papermc.io" run_pf)"
 expect "unreachable download hosts are named together" "cannot reach github.com fill-data.papermc.io over HTTPS" "$out"
+# The VM's rerun after an upgrade: one probe of fill-data.papermc.io timed out, a second a
+# minute later answered in under a second, and preflight had refused the install.
+rm -f "$pfroot"/tries.*
+out="$(PF_FLAKY=fill-data.papermc.io run_pf)"
+expect "a download host that drops one probe is tried again" "WENT ON" "$out"
+expect "two seconds later" "SLEEP: 2" "$out"
+[ "$(cat "$pfroot/tries.fill-data.papermc.io")" = xx ] && echo "PASS and not again once it answered" \
+  || { echo "FAIL: a host that answered its second probe was probed $(wc -c < "$pfroot/tries.fill-data.papermc.io") times"; fails=$((fails + 1)); }
+out="$(PF_STALL=fill-data.papermc.io run_pf)"
+expect "a host that completes TLS and then sits on the request is reachable" "WENT ON" "$out"
+rm -f "$pfroot"/tries.*
+out="$(PF_DOWN=fill-data.papermc.io run_pf)"
+expect "a host that never answers is refused" "cannot reach fill-data.papermc.io over HTTPS" "$out"
+[ "$(cat "$pfroot/tries.fill-data.papermc.io")" = xxx ] && echo "PASS after three probes" \
+  || { echo "FAIL: a host that never answered was probed $(wc -c < "$pfroot/tries.fill-data.papermc.io") times"; fails=$((fails + 1)); }
 
 # An install from a release's assets builds nothing: no Docker cache under /var/lib/containerd,
 # and Docker Hub only as the fallback for an image the release cannot supply. The downloaded

docs/operations.md

@@ -60,9 +60,10 @@ once, then stops with nothing touched **[SH-TESTED]**:
   (a Docker network there is the usual case); a wider route such as a `10.0.0.0/8` VPN
   is a warning;
 - HTTPS to the hosts it downloads from: GitHub and PaperMC's download API always, Docker
-  Hub when it builds images on the host. Installing a release, an unreachable Docker Hub
-  is a warning (it is needed only if an asset turns out unusable); from
-  `FELIS_ARTIFACT_DIR` it is not checked.
+  Hub when it builds images on the host. A host counts as reachable once a TLS handshake
+  with it completes, and each gets three tries two seconds apart. Installing a release, an
+  unreachable Docker Hub is a warning (it is needed only if an asset turns out unusable);
+  from `FELIS_ARTIFACT_DIR` it is not checked.
 
 `FELIS_PREFLIGHT=warn` reports the same problems as warnings and installs anyway, for a
 host the checks misjudge.