Why? pg_tblspc is a dir with nothing but symlinks the previous check on DT_LNK should work

You’re right. Hard-coding pg_tblspc was not needed. The real issue is avoiding symlink traversal during decrypt. I will update this to skip DT_LNK entries directly, so we no longer need the pg_tblspc name check.

Why removing DT_LNK?

Ohh sorry, removing DT_LNK was too broad. I will change it to keep symlink handling explicit, but skip symlink traversal before recursing. I think that will then prevents duplicate decrypt work through pg_tblspc/<oid>.

Is it because the untar process isn't reading the file path from pg_tblspc dir correclty and causing manifest verification failure? Can you giev an example?

Yes, this is for manifest verification. Tablespace tar entries are relative, for example:
PG_17_202406281/16385/16390

But the backup manifest stores the same file as:

pg_tblspc/<oid>/PG_17_202406281/16385/16390

So without prefixing pg_tblspc/<oid>, checksum/size verification uses the wrong key.

Hmm, did test fail because of this? Nice catch. Something is wrong with this, this looks like a regression issue.

Yes, the PG14 failure matched this path. In PG14 multiple tar streams. The old code freed/reset msg after one stream, and the next stream could hit bad state. The log showed Connection reset by peer, then the next test failed because the server stayed active.

This is a bit too ad hoc. I think we can reuse the same msg throughout the entire process. So at the end of out most iteration we should just do memset(msg, 0, ...) instead of pgmoneta_free_message(msg);. That'll save us some malloc.

Agreed. I will update this to reuse the same msg and just call:

memset(msg, 0, sizeof(struct message));

at the end of each tar iteration instead of freeing/reallocating it.

Does this have to be run time? Can we put this into Dockerfile? And should we?

I think Dockerfile alone won’t help here because /conf is bind-mounted from $PGCONF_DIRECTORY, so image-time permissions are hidden by the host mount. This cleanup handles stale host-side tablespace dirs from previous runs, where files may be owned by the container postgres user and block rm -Rf "$PGMONETA_ROOT_DIR". We can keep creation in run-postgresql, but cleanup still needs to happen in check.sh. What are your suggestions ?

I wonder if it's better to move the test cases to test_backup and test_restore, respectively.

That’s a valid option. I kept them in test_tablespaces.c because the tests share tablespace setup, symlink checks, and cleanup, and it lets us run only this coverage with -m tablespaces. I can move backup checks to test_backup.c and restore checks to test_restore.c if you want . Let me know :)