Re: recap of 9p problems in 6.9

["Eric Van Hensbergen" <eric.vanhensbergen@linux.dev>]

April 8, 2024 at 7:11 PM, "Eric Van Hensbergen" <eric.vanhensbergen@linux.dev> wrote:
> April 7, 2024 at 11:17 PM, "Dominique Martinet" <asmadeus@codewreck.org> wrote:
> >  * open / unlink / fstat|ftruncate etc fail

Okay - a lot more digging on this one.  I'm wondering if these issues have to do with changes inside vfs_lookup -- there as a bit of code that was in there that I didn't quite understand.  Previously we always created a new inode on lookup (unless we were in cache mode) -- this leads to a lot of duplicate inode structs floating around.  One of the simplification changes was to remove this (but I left the comment in which talks about parallel unlinks) -- now the unlink in this case isn't in parallel here but I wonder if we were sidestepping this bug because unlink was essentially operating on its own inode and the open fd had a distinct inode structure that wasn't perturbed.

There's potentailly a relationship to the ACL issues Kent observed as well since there's a similar code path in create which uses a new inode and then associates ACLs with that.

I'm just still struggling why this requires a new inode structure if there is one already associated with the qid.path -- unless the inode numbers as being stored in the qid.path are being reused too quickly and there are unlinked inodes which still maintain that number in the cache??

      -eric