On Sun, Mar 10, 2024 at 6:48 PM Thomas Munro <thomas.mu...@gmail.com> wrote: > I won't be surprised if the answer is: if you're holding a reference, > you have to get a pin (referring to bulk_write.c).
Ahhh, on second thoughts, I take that back, I think the original theory still actually works just fine. It's just that somewhere in our refactoring of that commit, when we were vacillating between different semantics for 'destroy' and 'release', I think we made a mistake: in RelationCacheInvalidate() I think we should now call smgrreleaseall(), not smgrdestroyall(). That satisfies the requirements for sinval queue overflow: we close md.c segments (and most importantly virtual file descriptors), so our lack of sinval records can't hurt us, we'll reopen all files as required. That's what CLOBBER_CACHE_ALWAYS is effectively testing (and more). But the smgr pointer remains valid, and retains only its "identity", eg hash table key, and that's also fine. It won't be destroyed until after the end of the transaction. Which was the point, and it allows things like bulk_write.c (and streaming_read.c) to hold an smgr reference. Right?