On Tue, Oct 31, 2023 at 06:12:26AM +0000, Al Viro wrote:
> On Tue, Oct 31, 2023 at 01:53:51AM +0000, Al Viro wrote:
>
> > Carving that series up will be interesting, though...
>
> I think I have a sane carve-up; will post if it survives testing.
OK, current variant survives local testing. Lives in
git://git.kernel.org/pub/scm/linux/kernel/git/viro/vfs.git #work.dcache
Shortlog:
fast_dput(): having ->d_delete() is not reason to delay refcount decrement
fast_dput(): handle underflows gracefully
fast_dput(): new rules for refcount
__dput_to_list(): do decrement of refcount in the caller
retain_dentry(): lift decrement of ->d_count into callers
__dentry_kill(): get consistent rules for ->d_count
dentry_kill(): don't bother with retain_dentry() on slow path
Call retain_dentry() with refcount 0
fold the call of retain_dentry() into fast_dput()
don't try to cut corners in shrink_lock_dentry()
fold dentry_kill() into dput()
get rid of __dget()
shrink_dentry_list(): no need to check that dentry refcount is marked dead
to_shrink_list(): call only if refcount is 0
switch select_collect{,2}() to use of to_shrink_list()
Diffstat:
fs/dcache.c | 268 ++++++++++++++++++++++--------------------------------------
1 file changed, 96 insertions(+), 172 deletions(-)
Individual patches in followups. Review and testing would be welcome,
and it's obviously next cycle fodder. Massage of refcounting is in the
first 11 commits, last 4 are followups.
