wip: support for quoting devices

man/ostree-commit.xml

@@ -177,6 +177,15 @@ License along with this library. If not, see <https://www.gnu.org/licenses/>.
                 </para></listitem>
             </varlistentry>
 
+            <varlistentry>
+                <term><option>--quote-devices</option></term>
+                <listitem><para>
+                    By default, ostree rejects block and character devices. This option instead "quotes" them
+                    as regular files. In order to be processed back into block and character devices,
+                    the corresponding <literal>--unquote-devices</literal> must be passed to <literal>ostree checkout</literal>.
+                </para></listitem>
+            </varlistentry>
+
             <varlistentry>
                 <term><option>--no-xattrs</option></term>
                 <listitem><para>

src/libostree/ostree-core-private.h

@@ -78,6 +78,20 @@ G_BEGIN_DECLS
  */
 #define _OSTREE_ZLIB_FILE_HEADER_GVARIANT_FORMAT G_VARIANT_TYPE ("(tuuuusa(ayay))")
 
+// ostree doesn't have native support for devices. Whiteouts in overlayfs
+// are a 0:0 character device, and in some cases people are copying docker/podman
+// style overlayfs container storage directly into ostree commits. This
+// adds special support for "quoting" the whiteout so it just appears as a regular
+// file in the ostree commit, but can be converted back into a character device
+// on checkout.
+#define OSTREE_QUOTED_OVERLAYFS_WHITEOUT_PREFIX ".ostree-wh."
+// Filename prefix to signify a character or block device. This
+// is not supported natively by ostree (because there is no reason
+// to ship devices in images). But because OCI supports it, and in
+// some cases one wants to map OCI to ostree, we have support for
+// "quoting" them.
+#define OSTREE_QUOTED_DEVICE_PREFIX ".ostree-quoted-device."
+
 GBytes *_ostree_file_header_new (GFileInfo *file_info, GVariant *xattrs);
 
 GBytes *_ostree_zlib_file_header_new (GFileInfo *file_info, GVariant *xattrs);
@@ -92,6 +106,9 @@ gboolean _ostree_stbuf_equal (struct stat *stbuf_a, struct stat *stbuf_b);
 GFileInfo *_ostree_mode_uidgid_to_gfileinfo (mode_t mode, uid_t uid, gid_t gid);
 gboolean _ostree_validate_structureof_xattrs (GVariant *xattrs, GError **error);
 
+gboolean _ostree_parse_quoted_device (const char *name, guint32 src_mode, const char **out_name, guint32 *out_mode,
+                                      dev_t *out_dev, GError **error);
+
 static inline void
 _ostree_checksum_inplace_from_bytes_v (GVariant *csum_v, char *buf)
 {

src/libostree/ostree-core.c

@@ -33,6 +33,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <sys/types.h>
+#include <sys/sysmacros.h>
 
 /* Generic ABI checks */
 G_STATIC_ASSERT (OSTREE_REPO_MODE_BARE == 0);
@@ -2331,6 +2332,70 @@ ostree_validate_structureof_dirmeta (GVariant *dirmeta, GError **error)
   return TRUE;
 }
 
+gboolean
+_ostree_parse_quoted_device (const char *name, guint32 src_mode, const char **out_name, guint32 *out_mode, dev_t *out_dev,
+                             GError **error)
+{
+  // Ensure we start with the quoted device prefix
+  const char *s = name;
+  const char *p = strchr (s, '.');
+  if (!p)
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  if (strncmp (s, OSTREE_QUOTED_DEVICE_PREFIX, p - name) != 0)
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  s += strlen (OSTREE_QUOTED_DEVICE_PREFIX);
+  g_assert (out_name);
+  *out_name = s;
+
+  // The input mode is the same as the source, but without the format bits
+  guint32 ret_mode = (src_mode & ~S_IFMT);
+
+  // Parse the mode
+  s++;
+  switch (*s)
+    {
+    case 'b':
+      ret_mode |= S_IFBLK;
+      break;
+    case 'c':
+      ret_mode |= S_IFCHR;
+      break;
+    case 'p':
+      ret_mode |= S_IFIFO;
+      break;
+    default:
+      return glnx_throw (error, "Invalid quoted device: %s", name);
+    }
+  s++;
+  if (*s != '.')
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  s++;
+  s = strchr (s, '.');
+  if (!s)
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  s++;
+  char *endptr;
+  unsigned int major, minor;
+  major = (unsigned int)g_ascii_strtoull (s, &endptr, 10);
+  if (errno == ERANGE)
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  s = endptr;
+  if (*s != '.')
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  s++;
+  minor = (unsigned int)g_ascii_strtoull (s, &endptr, 10);
+  if (errno == ERANGE)
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  g_assert (endptr);
+  if (*endptr != '\0')
+    return glnx_throw (error, "Invalid quoted device: %s", name);
+  g_assert (ret_mode);
+  *out_mode = ret_mode;
+  g_assert (out_dev);
+  *out_dev = makedev (major, minor);
+  return TRUE;
+}
+
 /**
  * ostree_commit_get_parent:
  * @commit_variant: Variant of type %OSTREE_OBJECT_TYPE_COMMIT

src/libostree/ostree-repo-checkout.c

@@ -35,14 +35,6 @@
 #define WHITEOUT_PREFIX ".wh."
 #define OPAQUE_WHITEOUT_NAME ".wh..wh..opq"
 
-// ostree doesn't have native support for devices. Whiteouts in overlayfs
-// are a 0:0 character device, and in some cases people are copying docker/podman
-// style overlayfs container storage directly into ostree commits. This
-// adds special support for "quoting" the whiteout so it just appears as a regular
-// file in the ostree commit, but can be converted back into a character device
-// on checkout.
-#define OSTREE_QUOTED_OVERLAYFS_WHITEOUT_PREFIX ".ostree-wh."
-
 /* Per-checkout call state/caching */
 typedef struct
 {
@@ -716,6 +708,9 @@ checkout_one_file_at (OstreeRepo *repo, OstreeRepoCheckoutAtOptions *options, Ch
   const gboolean is_unreadable = (!is_symlink && (source_mode & S_IRUSR) == 0);
   const gboolean is_whiteout = (!is_symlink && options->process_whiteouts
                                 && g_str_has_prefix (destination_name, WHITEOUT_PREFIX));
+  const gboolean is_quoted_device
+      = (!is_symlink && options->unquote_devices
+         && g_str_has_prefix (destination_name, OSTREE_QUOTED_DEVICE_PREFIX));
   const gboolean is_overlayfs_whiteout
       = (!is_symlink
          && g_str_has_prefix (destination_name, OSTREE_QUOTED_OVERLAYFS_WHITEOUT_PREFIX));
@@ -740,6 +735,16 @@ checkout_one_file_at (OstreeRepo *repo, OstreeRepoCheckoutAtOptions *options, Ch
 
       need_copy = FALSE;
     }
+  else if (is_quoted_device)
+    {
+      const char *devname;
+      dev_t dev;
+      guint32 mode;
+      if (!_ostree_parse_quoted_device (destination_name, source_mode, &devname, &mode, &dev, error))
+        return FALSE;
+      if (mknodat (destination_dfd, devname, (mode_t)mode, dev) < 0)
+        return glnx_throw_errno_prefix (error, "mknodat");
+    }
   else if (is_overlayfs_whiteout && options->process_passthrough_whiteouts)
     {
       const char *name = destination_name + (sizeof (OSTREE_QUOTED_OVERLAYFS_WHITEOUT_PREFIX) - 1);
@@ -1437,6 +1442,9 @@ canonicalize_options (OstreeRepo *self, OstreeRepoCheckoutAtOptions *options)
   /* Force USER mode for BARE_USER_ONLY always - nothing else makes sense */
   if (ostree_repo_get_mode (self) == OSTREE_REPO_MODE_BARE_USER_ONLY)
     options->mode = OSTREE_REPO_CHECKOUT_MODE_USER;
+
+  if (options->unquote_devices)
+    options->process_whiteouts = TRUE;
 }
 
 /**

src/libostree/ostree-repo-commit.c

@@ -3450,6 +3450,202 @@ write_dir_entry_to_mtree_internal (OstreeRepo *self, OstreeRepoFile *repo_dir,
   return TRUE;
 }
 
+static gboolean
+write_quoted_device (OstreeRepo *self, OstreeRepoFile *repo_dir,
+                                 GFileEnumerator *dir_enum, GLnxDirFdIterator *dfd_iter,
+                                 WriteDirContentFlags writeflags, GFileInfo *child_info,
+                                 OstreeMutableTree *mtree, OstreeRepoCommitModifier *modifier,
+                                 GPtrArray *path, GCancellable *cancellable, GError **error)
+{
+  g_assert (dir_enum != NULL || dfd_iter != NULL);
+
+  GFileType file_type = g_file_info_get_file_type (child_info);
+  const char *name = g_file_info_get_name (child_info);
+
+  /* Load flags into boolean constants for ease of readability (we also need to
+   * NULL-check modifier)
+   */
+  const gboolean canonical_permissions
+      = self->mode == OSTREE_REPO_MODE_BARE_USER_ONLY
+        || (modifier
+            && (modifier->flags & OSTREE_REPO_COMMIT_MODIFIER_FLAGS_CANONICAL_PERMISSIONS));
+  const gboolean devino_canonical
+      = modifier && (modifier->flags & OSTREE_REPO_COMMIT_MODIFIER_FLAGS_DEVINO_CANONICAL);
+  /* We currently only honor the CONSUME flag in the dfd_iter case to avoid even
+   * more complexity in this function, and it'd mostly only be useful when
+   * operating on local filesystems anyways.
+   */
+  const gboolean delete_after_commit
+      = dfd_iter && modifier && (modifier->flags & OSTREE_REPO_COMMIT_MODIFIER_FLAGS_CONSUME);
+
+  /* Build the full path which we need for callbacks */
+  g_ptr_array_add (path, (char *)name);
+  g_autofree char *child_relpath = ptrarray_path_join (path);
+
+  /* Call the filter */
+  g_autoptr (GFileInfo) modified_info = NULL;
+  OstreeRepoCommitFilterResult filter_result = _ostree_repo_commit_modifier_apply (
+      self, modifier, child_relpath, child_info, &modified_info);
+  const gboolean child_info_was_modified = !_ostree_gfileinfo_equal (child_info, modified_info);
+
+  if (filter_result != OSTREE_REPO_COMMIT_FILTER_ALLOW)
+    {
+      g_ptr_array_remove_index (path, path->len - 1);
+      if (delete_after_commit)
+        {
+          g_assert (dfd_iter);
+          if (!glnx_shutil_rm_rf_at (dfd_iter->fd, name, cancellable, error))
+            return FALSE;
+        }
+      /* Note: early return */
+      return TRUE;
+    }
+
+  guint32 src_mode = g_file_info_get_attribute_uint32 (src_info, "unix::mode")';'
+  switch (file_type)
+    {
+    case G_FILE_TYPE_SYMBOLIC_LINK:
+    case G_FILE_TYPE_REGULAR:
+      break;
+    default:
+      return glnx_throw (error, "Unsupported file type for file: '%s'", child_relpath);
+    }
+
+  g_autoptr (GFile) child = NULL;
+  if (dir_enum != NULL)
+    child = g_file_enumerator_get_child (dir_enum, child_info);
+
+  /* Our filters have passed, etc.; now we prepare to write the content object */
+  glnx_autofd int file_input_fd = -1;
+
+  /* Open the file now, since it's better for reading xattrs
+   * rather than using the /proc/self/fd links.
+   *
+   * TODO: Do this lazily, since for e.g. bare-user-only repos
+   * we don't have xattrs and don't need to open every file
+   * for things that have devino cache hits.
+   */
+  if (file_type == G_FILE_TYPE_REGULAR && dfd_iter != NULL)
+    {
+      if (!glnx_openat_rdonly (dfd_iter->fd, name, FALSE, &file_input_fd, error))
+        return FALSE;
+    }
+
+  g_autoptr (GVariant) xattrs = NULL;
+  gboolean xattrs_were_modified;
+  if (dir_enum != NULL)
+    {
+      if (!get_final_xattrs (self, modifier, child_relpath, child_info, child, -1, name,
+                             source_xattrs, &xattrs, &xattrs_were_modified, cancellable, error))
+        return FALSE;
+    }
+  else
+    {
+      /* These contortions are basically so we use glnx_fd_get_all_xattrs()
+       * for regfiles, and glnx_dfd_name_get_all_xattrs() for symlinks.
+       */
+      int xattr_fd_arg = (file_input_fd != -1) ? file_input_fd : dfd_iter->fd;
+      const char *xattr_path_arg = (file_input_fd != -1) ? NULL : name;
+      if (!get_final_xattrs (self, modifier, child_relpath, child_info, child, xattr_fd_arg,
+                             xattr_path_arg, source_xattrs, &xattrs, &xattrs_were_modified,
+                             cancellable, error))
+        return FALSE;
+    }
+
+  /* Used below to see whether we can do a fast path commit */
+  const gboolean modified_file_meta = child_info_was_modified || xattrs_were_modified;
+
+  /* A big prerequisite list of conditions for whether or not we can
+   * "adopt", i.e. just checksum and rename() into place
+   */
+  const gboolean can_adopt_basic = file_type == G_FILE_TYPE_REGULAR && dfd_iter != NULL
+                                   && delete_after_commit
+                                   && ((writeflags & WRITE_DIR_CONTENT_FLAGS_CAN_ADOPT) > 0);
+  gboolean can_adopt = can_adopt_basic;
+  /* If basic prerquisites are met, check repo mode specific ones */
+  if (can_adopt)
+    {
+      /* For bare repos, we could actually chown/reset the xattrs, but let's
+       * do the basic optimizations here first.
+       */
+      if (self->mode == OSTREE_REPO_MODE_BARE)
+        can_adopt = !modified_file_meta;
+      else if (self->mode == OSTREE_REPO_MODE_BARE_USER_ONLY)
+        can_adopt = canonical_permissions;
+      else
+        /* This covers bare-user and archive.  See comments in adopt_and_commit_regfile()
+         * for notes on adding bare-user later here.
+         */
+        can_adopt = FALSE;
+    }
+  gboolean did_adopt = FALSE;
+
+  /* The very fast path - we have a devino cache hit, nothing to write */
+  if (loose_checksum && !modified_file_meta)
+    {
+      if (!ostree_mutable_tree_replace_file (mtree, name, loose_checksum, error))
+        return FALSE;
+
+      g_mutex_lock (&self->txn_lock);
+      self->txn.stats.devino_cache_hits++;
+      g_mutex_unlock (&self->txn_lock);
+    }
+  /* Next fast path - we can "adopt" the file */
+  else if (can_adopt)
+    {
+      char checksum[OSTREE_SHA256_STRING_LEN + 1];
+      if (!adopt_and_commit_regfile (self, dfd_iter->fd, name, modified_info, xattrs, checksum,
+                                     cancellable, error))
+        return FALSE;
+      if (!ostree_mutable_tree_replace_file (mtree, name, checksum, error))
+        return FALSE;
+      did_adopt = TRUE;
+    }
+  else
+    {
+      g_autoptr (GInputStream) file_input = NULL;
+
+      if (file_type == G_FILE_TYPE_REGULAR)
+        {
+          if (dir_enum != NULL)
+            {
+              g_assert (child != NULL);
+              file_input = (GInputStream *)g_file_read (child, cancellable, error);
+              if (!file_input)
+                return FALSE;
+            }
+          else
+            {
+              /* We already opened the fd above */
+              file_input = g_unix_input_stream_new (file_input_fd, FALSE);
+            }
+        }
+
+      g_autofree guchar *child_file_csum = NULL;
+      if (!write_content_object (self, NULL, file_input, modified_info, xattrs, &child_file_csum,
+                                 cancellable, error))
+        return FALSE;
+
+      char tmp_checksum[OSTREE_SHA256_STRING_LEN + 1];
+      ostree_checksum_inplace_from_bytes (child_file_csum, tmp_checksum);
+      if (!ostree_mutable_tree_replace_file (mtree, name, tmp_checksum, error))
+        return FALSE;
+    }
+
+  /* Process delete_after_commit. In the adoption case though, we already
+   * took ownership of the file above, usually via a renameat().
+   */
+  if (delete_after_commit && !did_adopt)
+    {
+      if (!glnx_unlinkat (dfd_iter->fd, name, 0, error))
+        return FALSE;
+    }
+
+  g_ptr_array_remove_index (path, path->len - 1);
+
+  return TRUE;
+}
+
 /* Given either a dir_enum or a dfd_iter, writes a non-dir (regfile/symlink) to
  * the mtree.
  */
@@ -3889,6 +4085,14 @@ write_dfd_iter_to_mtree_internal (OstreeRepo *self, GLnxDirFdIterator *src_dfd_i
                                          error))
             return FALSE;
         }
+      else if (modifier->flags & OSTREE_REPO_COMMIT_MODIFIER_FLAGS_QUOTE_DEVICES)
+        {
+          if (!write_quoted_device (self, NULL, NULL, src_dfd_iter, flags, child_info, mtree,
+                                    modifier, path, cancellable, error))
+            return FALSE;
+          // Note we skip over the code below
+          continue;
+        }
       else
         {
           return glnx_throw (error, "Not a regular file or symlink: %s", dent->d_name);