Fix LFS timeouts when trying to save large files

The following was happening:

1. Workhorse stores an LFS file in /var/opt/gitlab/gitlab-rails/shared/lfs-objects
2. CarrierWave then renames the file to a temporary directory (e.g. /opt/gitlab/embedded/service/gitlab-rails/tmp)
3. CarrierWave then renames the file to its final location (e.g. /var/opt/gitlab/gitlab-rails/shared/lfs-objects)

When the LFS upload path was on a different filesystem than the Rails
installation, step 2 could take a longer than 10 seconds, at which point Workhorse
would time out with "badgateway: failed after 10s: context canceled".

This change makes the work path in the same root as the LFS storage path, preventing moves
across filesystems.

Closes #33218

1 files changed, 16 insertions, 0 deletions

diff --git a/app/uploaders/lfs_object_uploader.rb b/app/uploaders/lfs_object_uploader.rb
index 95a891111e1..02589959c2f 100644
--- a/app/uploaders/lfs_object_uploader.rb
+++ b/app/uploaders/lfs_object_uploader.rb
@@ -12,4 +12,20 @@ class LfsObjectUploader < GitlabUploader
   def filename
     model.oid[4..-1]
   end
+
+  def work_dir
+    File.join(Gitlab.config.lfs.storage_path, 'tmp', 'work')
+  end
+
+  private
+
+  # To prevent LFS files from moving across filesystems, override the default
+  # implementation:
+  # http://github.com/carrierwaveuploader/carrierwave/blob/v1.0.0/lib/carrierwave/uploader/cache.rb#L181-L183
+  def workfile_path(for_file = original_filename)
+    # To be safe, keep this directory outside of the the cache directory
+    # because calling CarrierWave.clean_cache_files! will remove any files in
+    # the cache directory.
+    File.join(work_dir, @cache_id, version_name.to_s, for_file)
+  end
 end