fix(parity): rsync 3.4.1 symlink and special-node semantics (#287, #288)

#287: - --safe-links: keep safe in-tree links AS symlinks and skip unsafe (absolute or ".."-escaping) ones, mirroring rsync's unsafe_symlink(). Skipped links are recorded as delete-protected so --delete does not remove their destination mirror (no silent data loss). - --copy-unsafe-links: preserve safe links as symlinks and dereference only unsafe ones. - --munge-links: receiver-side rewrite storing /rsyncd-munged/-prefixed targets (rsync parity), replacing the no-op #SYMLINK sender prefix. - -l: store the target verbatim, including absolute and ".." targets (rsync -l parity); the old receiver containment silently dropped them. #288: - --specials: recreate unix-domain sockets via mknod(S_IFSOCK), which Linux permits unprivileged; keep EEXIST/EPERM skip behavior. - --copy-devices: copy a device's content into a regular file when requested; skip unrequested non-regular entries like rsync's default.
2026-09-15 21:57:48 +02:00
parent 23552e823d
commit ea4ab661b4
8 changed files with 573 additions and 250 deletions
@@ -415,10 +415,69 @@ bool file_get_trust_sender(void) {
  return file_trust_sender;
 }

-/* True when `target` is a lexical symlink target that can never escape the
- * receive root once created beneath it: relative (not absolute) and containing
- * no ".." path component.  Used by --munge-links' sender-side containment: an
- * escaping target is never transmitted (the entry is skipped/contained). */
+/* rsync 3.4.1 unsafe_symlink(): true when `target` (the link's destination
+ * string) points outside the transfer tree rooted at the symlink's own
+ * location.  `link_path` is the symlink's path relative to the top of the
+ * transfer (including its name).  This is a purely lexical test matching
+ * rsync's util1.c: absolute/empty targets are always unsafe; leading "../"
+ * components are counted against the symlink's own directory depth; a ".."
+ * that would climb above the transfer root is unsafe.  rsync 3.4.1 additionally
+ * rejects any INTERNAL "/../" component and a trailing "/..". */
+bool file_symlink_unsafe(const char* target, const char* link_path) {
+  if (!target || target[0] == '\0' || target[0] == '/')
+    return true;
+  const char* rest = target;
+  while (strncmp(rest, "../", 3) == 0) {
+    rest += 3;
+    while (*rest == '/')
+      rest++;
+  }
+  if (strstr(rest, "/../") != NULL)
+    return true;
+  size_t target_len = strlen(target);
+  if (target_len > 3 && strcmp(&target[target_len - 3], "/..") == 0)
+    return true;
+
+  int depth = 0;
+  const char* name;
+  const char* slash;
+  const char* src = link_path ? link_path : "";
+  for (name = src; (slash = strchr(name, '/')) != NULL; name = slash + 1) {
+    if (*name == '.' && (name[1] == '/' || (name[1] == '.' && name[2] == '/'))) {
+      if (name[1] == '.')
+        depth = 0;
+    } else {
+      depth++;
+    }
+    while (slash[1] == '/')
+      slash++;
+  }
+  if (*name == '.' && name[1] == '.' && name[2] == '\0')
+    depth = 0;
+
+  for (name = target; (slash = strchr(name, '/')) != NULL; name = slash + 1) {
+    if (*name == '.' && (name[1] == '/' || (name[1] == '.' && name[2] == '/'))) {
+      if (name[1] == '.') {
+        if (--depth < 0)
+          return true;
+      }
+    } else {
+      depth++;
+    }
+    while (slash[1] == '/')
+      slash++;
+  }
+  if (*name == '.' && name[1] == '.' && name[2] == '\0')
+    depth--;
+  return depth < 0;
+}
+
+/* Strict lexical helper: true when `target` is relative (not absolute) and
+ * contains no ".." component at all, so it can never escape the directory it
+ * is created in.  This is stricter than rsync's unsafe_symlink() (which allows
+ * an in-tree ".."); the scanner/receiver use file_symlink_unsafe()/--safe-links
+ * for rsync parity, and this helper is retained for callers that want the
+ * ".."-free guarantee. */
 bool file_symlink_target_contained(const char* target) {
  if (!target || target[0] == '\0' || target[0] == '/')
    return false;
@@ -449,8 +508,9 @@ bool file_symlink_unmunge(char* target) {
  return true;
 }

-/* Owned copy of `target` prefixed with SYMLINK_MUNGE_PREFIX (the sender-side
- * --munge-links rewriting).  Returns NULL on allocation failure. */
+/* Owned copy of `target` prefixed with SYMLINK_MUNGE_PREFIX (the receiver-side
+ * --munge-links rewriting, matching rsync's receiver).  Returns NULL on
+ * allocation failure. */
 char* file_symlink_munge(const char* target) {
  if (!target)
    return NULL;
@@ -471,23 +531,14 @@ char* file_symlink_munge(const char* target) {
 * the target is ever followed.  The final component is never dereferenced: an
 * existing non-directory entry at `path` is unlinked by name before the link is
 * placed; an existing directory there is left untouched (returns false, so a
- * caller can treat it as a collision).  As a receiver-side trust-boundary
- * invariant, `target` must be file_symlink_target_contained() (relative and
- * ".."-free): an absolute or escaping target is rejected outright (returns
- * false) so a malicious sender can never materialize a symlink that points
- * outside the receive root. */
+ * caller can treat it as a collision).  The link VALUE `target` is copied
+ * verbatim, matching rsync -l (which stores absolute and ".."-bearing targets
+ * as-is); target policy is the caller's job -- the scanner applies
+ * --safe-links/--copy-unsafe-links, and the receiver applies --munge-links.
+ * The PLACEMENT path is always confined below the authorized root. */
 bool file_symlink_at_secure(const char* path, const char* target) {
-  /* The link itself (`path`) is always kept below the authorized root.  The
-     TARGET may point anywhere: normally only a contained (relative, ".."-free)
-     target is permitted so a malicious sender can never plant a symlink that
-     later dereferences outside the root.  Under --trust-sender that target
-     containment check is relaxed (the receiver trusts the sender and copies the
-     link verbatim, matching rsync -l), but path/leaf confinement is never
-     disabled, so the link still cannot be placed outside the tree. */
  if (!path || !target || has_path_traversal(path))
    return false;
-  if (!file_trust_sender && !file_symlink_target_contained(target))
-    return false;
  char* leaf = NULL;
  int parent_fd = file_open_secure_parent(path, &leaf, true);
  if (parent_fd < 0)
@@ -44,12 +44,20 @@ int file_open_for_read(const char* path);
 bool file_write_to_disk(const char* path, const void* data, unsigned long long data_size,
                        bool inplace, bool sparse);

-/* Symlink trust-boundary helpers (Phase 4, symlink wave).  --munge-links
- * sender-side marker: every transmitted symlink target is prefixed with this
- * while the flag is on; the receiver strips it to restore the real target. */
-#define SYMLINK_MUNGE_PREFIX "#SYMLINK/"
+/* Symlink trust-boundary helpers (Phase 4, symlink wave; rsync parity).
+ * --munge-links is a RECEIVER-side rewrite: rsync prefixes every stored symlink
+ * target with this marker, making the link unusable while the referenced
+ * directory does not exist.  A SENDER receiving a munged source strips it back
+ * off before transmitting (so a munged tree round-trips through the receiver's
+ * re-munging). */
+#define SYMLINK_MUNGE_PREFIX "/rsyncd-munged/"

 char* file_symlink_munge(const char* target);
+/* rsync 3.4.1 unsafe_symlink(): true when `target` escapes the transfer tree
+ * rooted at `link_path` (the symlink's transfer-relative path incl. its name).
+ * Absolute/empty targets and targets climbing above the transfer root (via
+ * "..") are unsafe, as are internal "/../" components and trailing "/..". */
+bool file_symlink_unsafe(const char* target, const char* link_path);
 /* True when a lexical target is relative and contains no ".." component, so it
 * can never escape the receive root once created beneath it. */
 bool file_symlink_target_contained(const char* target);
@@ -57,8 +65,9 @@ bool file_symlink_target_contained(const char* target);
 * returns true when a marker was removed. */
 bool file_symlink_unmunge(char* target);
 /* Create a symlink at `path` -> `target`, confined below the authorized root
- * (O_NOFOLLOW parent walk, symlinkat; the target is never followed).  Returns
- * false when a directory already occupies `path`. */
+ * (O_NOFOLLOW parent walk, symlinkat; the target is never followed).  The link
+ * value is copied verbatim (rsync -l); only the placement path is confined.
+ * Returns false when a directory already occupies `path`. */
 bool file_symlink_at_secure(const char* path, const char* target);
 /* --keep-dirlinks (-K) receiver process-wide policy: allow an in-root existing
 * symlink-to-directory to be followed as a directory. */
@@ -358,14 +358,6 @@ static FileSaveResult file_save_special_to_disk(const char* root_directory, cons
    log_message(LOG_LEVEL_ERROR, "Special node has no device/FIFO/socket mode");
    return FILE_SAVE_ERROR;
  }
-  if (is_sock) {
-    /* No standard filesystem call recreates a socket; best-effort unsupported. */
-    char* escaped_path = output_escape(file->path, log_get_8_bit_output());
-    log_message(LOG_LEVEL_WARNING, "socket not recreated: %s (unsupported; skipped)",
-                escaped_path ? escaped_path : "<allocation failed>");
-    free(escaped_path);
-    return FILE_SAVE_SKIPPED;
-  }
  if (is_char || is_blk) {
    if (!config || !config->preserve_devices)
      return FILE_SAVE_SKIPPED;
@@ -383,7 +375,10 @@ static FileSaveResult file_save_special_to_disk(const char* root_directory, cons
      free(escaped_path);
      return FILE_SAVE_SKIPPED;
    }
-  } else if (is_fifo) {
+  } else if (is_fifo || is_sock) {
+    /* FIFOs and unix sockets are recreated by --specials.  mknod(S_IFSOCK)
+       works unprivileged on Linux (the node carries no live socket), so unlike
+       a socket bound to a live fd it can be materialized. */
    if (!config || !config->preserve_specials)
      return FILE_SAVE_SKIPPED;
  }
@@ -433,9 +428,12 @@ static FileSaveResult file_save_special_to_disk(const char* root_directory, cons
  } else if (is_blk) {
    create_mode = S_IFBLK;
    rdev = makedev((unsigned)file->rdev_major, (unsigned)file->rdev_minor);
+  } else if (is_sock) {
+    create_mode = S_IFSOCK;
  } else {
    create_mode = S_IFIFO;
  }
+  const char* node_kind = (is_char || is_blk) ? "device" : (is_fifo ? "FIFO" : "socket");
  /* The creation permission bits come from the source only under -p/--perms;
   * otherwise a safe default (0644, group/other write never granted) keeps an
   * unprivileged no--p run from materializing a world-writable node. */
@@ -450,7 +448,7 @@ static FileSaveResult file_save_special_to_disk(const char* root_directory, cons
      struct stat st;
      if (fstatat(parent_fd, leaf, &st, AT_SYMLINK_NOFOLLOW) == 0 &&
          ((is_char && S_ISCHR(st.st_mode)) || (is_blk && S_ISBLK(st.st_mode)) ||
-           (is_fifo && S_ISFIFO(st.st_mode)))) {
+           (is_fifo && S_ISFIFO(st.st_mode)) || (is_sock && S_ISSOCK(st.st_mode)))) {
        close(parent_fd);
        free(leaf);
        free(destination);
@@ -458,7 +456,7 @@ static FileSaveResult file_save_special_to_disk(const char* root_directory, cons
      }
      char* escaped_path = output_escape(file->path, log_get_8_bit_output());
      log_message(LOG_LEVEL_WARNING, "refusing to replace existing entry with %s: %s (skipped)",
-                  is_fifo ? "FIFO" : "device", escaped_path ? escaped_path : "<allocation failed>");
+                  node_kind, escaped_path ? escaped_path : "<allocation failed>");
      free(escaped_path);
    } else if (errno == EPERM || errno == EACCES) {
      /* Missing CAP_MKNOD / parent write permission: the environment cannot
@@ -467,14 +465,12 @@ static FileSaveResult file_save_special_to_disk(const char* root_directory, cons
      log_message(LOG_LEVEL_WARNING,
                  "skipping %s: cannot create %s node (%s)\n"
                  "  --devices/--specials node creation needs privilege (CAP_MKNOD)",
-                  escaped_path ? escaped_path : "<allocation failed>", is_fifo ? "FIFO" : "device",
-                  strerror(errno));
+                  escaped_path ? escaped_path : "<allocation failed>", node_kind, strerror(errno));
      free(escaped_path);
    } else {
      char* escaped_path = output_escape(file->path, log_get_8_bit_output());
-      log_message(LOG_LEVEL_WARNING, "failed to create %s %s: %s (skipped)",
-                  is_fifo ? "FIFO" : "device", escaped_path ? escaped_path : "<allocation failed>",
-                  strerror(errno));
+      log_message(LOG_LEVEL_WARNING, "failed to create %s %s: %s (skipped)", node_kind,
+                  escaped_path ? escaped_path : "<allocation failed>", strerror(errno));
      free(escaped_path);
    }
    close(parent_fd);
@@ -710,26 +706,23 @@ FileSaveResult file_save_to_disk_full(const char* root_directory, const File* fi
    char* link_path = path_cat(root_directory, file->path);
    if (!link_path)
      return FILE_SAVE_ERROR;
-    /* Restore the real target by stripping the sender's --munge-links marker.
-       Only unmunge when the policy was negotiated: a plain -l run must preserve
-       a source symlink whose target genuinely begins with the marker verbatim. */
+    /* The link value is stored verbatim (rsync -l parity: absolute and
+       ".."-bearing targets are preserved; the scanner's --safe-links /
+       --copy-unsafe-links decide which links are sent at all).  --munge-links
+       is a RECEIVER-side rewrite: the stored target is prefixed with
+       /rsyncd-munged/, making the link unusable while the referenced directory
+       does not exist -- exactly as rsync's receiver munges.  Only the link's
+       own placement path is confined below the receive root. */
+    bool munge = config && config->munge_links;
    char* target = str_dup(file->symlink_target);
    bool ok = target != NULL;
-    if (ok && config && config->munge_links)
-      file_symlink_unmunge(target);
-    /* Receiver-side trust boundary (independent of the sender): a target that
-       could escape the receive root (absolute, or relative-with-"..") is never
-       materialized.  It is contained (the entry is skipped) rather than failing
-       the whole transfer, so a hostile sender can inject a broken symlink but
-       can never redirect it outside the root.  --trust-sender deliberately
-       relaxes this receiver-side re-validation: a trusted sender's escaping
-       symlink target is copied verbatim (rsync -l parity).  The low-level
-       leaf/destination confinement in file_symlink_at_secure still ensures the
-       link itself is placed inside the authorized root. */
-    if (ok && !file_get_trust_sender() && !file_symlink_target_contained(target))
-      ok = false;
+    if (ok && munge) {
+      char* munged = file_symlink_munge(target);
+      free(target);
+      target = munged;
+      ok = target != NULL;
+    }
    if (!ok) {
-      /* Skip the escaping/empty target (contained) rather than abort. */
      free(target);
      free(link_path);
      return FILE_SAVE_SKIPPED;