On Thu, Nov 21, 2024 at 11:12 AM Christian Brauner <brauner@xxxxxxxxxx> wrote:
> I think that i_devices should be moved into the union as it's really
> only used with i_cdev but it's not that easily done because list_head
> needs to be initialized. I roughly envisioned something like:
>
> union {
> struct {
> struct cdev *i_cdev;
> struct list_head i_devices;
> };
> struct {
> char *i_link;
> unsigned int i_link_len;
> };
> struct pipe_inode_info *i_pipe;
> unsigned i_dir_seq;
> };
>
> But it's not important enough imho.
I thought about doing that, but decided not to. I mentioned some time
ago that the layout of struct inode is false-sharing friendly and the
kernel is not in shape where this can be sensibly fixed yet and it may
or may not affect what to do with the above.
On the stock kernel the issues are:
- a spurious lockref acquire/release -- I posted a patch for it, Al
did not like it and wrote his own, does not look like it landed though
- apparmor -- everything serializes on label ref management (this *is*
used by ubuntu for example, but also the lkp machinery). Other people
posted patchsets to get rid of the problem, but they ran into their
own snags. I have a wip with of my own.
Anyhow, with these eliminated it will be possible to evaluate what
happens with inode rearrengements. Until then I think any messing with
the layout is best avoided.
thanks for taking the patchset
--
Mateusz Guzik <mjguzik gmail.com>
