Messages in this thread | | | Date | Mon, 16 Mar 2020 18:35:18 +0000 | From | Al Viro <> | Subject | Re: [RFC PATCH] locks:Remove spinlock in unshare_files |
| |
On Mon, Mar 16, 2020 at 02:39:16PM +0100, Peter Zijlstra wrote:
> > > diff --git a/kernel/fork.c b/kernel/fork.c > > > index 60a1295..fe54600 100644 > > > --- a/kernel/fork.c > > > +++ b/kernel/fork.c > > > @@ -3041,9 +3041,7 @@ int unshare_files(struct files_struct **displaced) > > > return error; > > > } > > > *displaced = task->files; > > > - task_lock(task); > > > - task->files = copy; > > > - task_unlock(task); > > > + WRITE_ONCE(task->files, copy); > > > return 0; > > > } > > AFAICT this is completely and utterly buggered. > > IFF task->files was lockless, like say RCU, then you'd still need > smp_store_release(). But if we look at fs/file.c then everything uses > task_lock() and removing it like the above is actively broken.
The problem is not fs/file.c; it's the code that does (read-only) access to *other* threads' ->files. procfs, SAK, some cgroup shite (pardon the redundancy)... All of those rely upon task_lock.
FWIW, having just grepped around, I'm worried about the crap io_uring is pulling off - interplay with unshare(2) could be unpleasant.
In any case - task_lock in the code that assigns to ->files (and it's not just unshare_files()) serves to protect the 3rd-party readers (including get_files_struct()) from having the fucker taken apart under them. It's not just freeing the thing - it's the entire close_files().
And no, we do *NOT* want to convert everything to get_files_struct() + being clever in it. I would rather have get_files_struct() taken out and shot, TBH - the only real reason it hadn't been killed years ago is the loop in proc_readfd_common()...
I'd prefer to have 3rd-party readers indicate their interest in a way that would be distinguishable from normal references, with close_files() waiting until all of those are gone. One way to do that would be * secondary counter in files_struct * rcu-delayed freeing of actual structure (not a problem) * rcu_read_lock in 3rd-party readers (among other things it means that proc_readfd_common() would need to be rearchitected a bit) * close_files() starting with subtraction of large constant from the secondary counter and then spinning until it gets to -<large constant> * 3rd-party readers (under rcu_read_lock()) fetching task->files, bumping the secondary counter unless it's negative, doing their thing, then decrementing the counter.
| |