The inline_data feature for ceph allowed you to store the first chunk of file data in the MDS instead of the OSD, using the cap system to keep things in sync. The kernel client implementation has always been a bit buggy and racy, and user uptake of this feature was pretty low. A couple of years ago, we decided to formally deprecate the inline_data feature entirely, and marked it as such in the Octopus release. We're still waiting on a few pieces in userland [1] before we can rip out support there so we need to keep support for inline_data in the client for now. This patch series changes the client from uninlining the data on the first write, to doing so at open time. This gets the uninlining out of the more fiddly write codepaths, and fits with the larger goal of aggressively deprecating inline_data. The original series was proposed by David. I added fixes for a few bugs found during testing and rolled in a related fix to change how we acquire the inode pointer. [1]: https://tracker.ceph.com/issues/52916 David Howells (2): ceph: Make ceph_netfs_issue_op() handle inlined data ceph: Uninline the data on a file opened for writing Jeff Layton (1): ceph: switch netfs read ops to use rreq->inode instead of rreq->mapping->host fs/ceph/addr.c | 239 +++++++++++++++++++++++------------------------- fs/ceph/file.c | 32 +++---- fs/ceph/super.h | 2 +- 3 files changed, 129 insertions(+), 144 deletions(-) -- 2.34.1
