Re: [PATCH] eventfs: Have inodes have unique inode numbers

[Steven Rostedt]

On Sun, 28 Jan 2024 12:53:31 -0800
Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:

> Honestly, you should just *always* do refcounting. No "free after RCU
> delay" as an alternative. Just refcount it.
> 
> Now, the RCU delay may be needed if the lookup of said structure
> happens under RCU, but no, saying "I use SRCU to make sure the
> lifetime is at least X" is just broken.
> 
> The refcount is what gives the lifetime. Any form of RCU-delaying
> should then be purely about non-refcounting RCU lookups that may
> happen as the thing is dying (and said lookup should *look* at the
> refcount and say "oh, this is dead, I'm not returning this".

The deleting of the ei is done outside the VFS logic. I use SRCU to
synchronize looking at the ei children in the lookup. On deletion, I
grab the eventfs_mutex, set ei->is_freed and then wait for SRCU to
finish before freeing.

The lookup checks ei->is_freed and doesn't do anything if set, but most
that logic is under the SRCU, which is what I want to make sure is
finished before the ei is deleted.

Hmm, I still need the logic for iput(), as dentry->d_fsdata can still
access the ei. That's where I need to have the ref counters. For a
lookup, I need to up the ref count when I create a new inode for the ei
or its children. Then in the iput() I decrement the ei ref count. I can
only free the ei if the ref count is zero.

The ref count is for knowing if an ei is referenced by a
dentry->d_fsdata, and the SRCU is to make sure there's no lookups
accessing an ei.

-- Steve