On Fri, Mar 07, 2025 at 05:11:55PM +0100, Mateusz Guzik wrote: > Atomics are only needed for a combination of io_uring and audit. > > Regular file access (even with audit) gets around fine without them. > > With this patch 'struct filename' starts with being refcounted using > regular ops. > > In order to avoid API explosion in the getname*() family, a dedicated > routine is added to switch the obj to use atomics. > > This leaves the room for merely issuing getname(), not issuing the > switch and still trying to manipulate the refcount from another thread. > > Catching such cases is facilitated by CONFIG_DEBUG_VFS-dependent > tracking of who created the given filename object and having refname() > and putname() detect if another thread is trying to modify them. Not a good way to handle that, IMO. Atomics do hurt there, but they are only plastering over the real problem - names formed in one thread, inserted into audit context there and operation involving them happening in a different thread. Refcounting avoids an instant memory corruption, but the real PITA is in audit users of that stuff. IMO we should *NOT* grab an audit names slot at getname() time - that ought to be done explicitly at later points. The obstacle is that currently there still are several retry loop with getname() done in it; I've most of that dealt with, need to finish that series. And yes, refcount becomes non-atomic as the result.
