On Thu, 07 Dec 2023, Christian Brauner wrote:
> On Tue, Dec 05, 2023 at 04:31:51PM -0700, Jens Axboe wrote:
> > On 12/5/23 4:23 PM, NeilBrown wrote:
> > > On Wed, 06 Dec 2023, NeilBrown wrote:
> > >> On Wed, 06 Dec 2023, Jens Axboe wrote:
> > >>> On 12/5/23 2:58 PM, Jens Axboe wrote:
> > >>>> On 12/5/23 2:28 PM, NeilBrown wrote:
> > >>>>> On Tue, 05 Dec 2023, Christian Brauner wrote:
> > >>>>>> On Mon, Dec 04, 2023 at 03:09:44PM -0700, Jens Axboe wrote:
> > >>>>>>> On 12/4/23 2:02 PM, NeilBrown wrote:
> > >>>>>>>> It isn't clear to me what _GPL is appropriate, but maybe the rules
> > >>>>>>>> changed since last I looked..... are there rules?
> > >>>>>>>>
> > >>>>>>>> My reasoning was that the call is effectively part of the user-space
> > >>>>>>>> ABI. A user-space process can call this trivially by invoking any
> > >>>>>>>> system call. The user-space ABI is explicitly a boundary which the GPL
> > >>>>>>>> does not cross. So it doesn't seem appropriate to prevent non-GPL
> > >>>>>>>> kernel code from doing something that non-GPL user-space code can
> > >>>>>>>> trivially do.
> > >>>>>>>
> > >>>>>>> By that reasoning, basically everything in the kernel should be non-GPL
> > >>>>>>> marked. And while task_work can get used by the application, it happens
> > >>>>>>> only indirectly or implicitly. So I don't think this reasoning is sound
> > >>>>>>> at all, it's not an exported ABI or API by itself.
> > >>>>>>>
> > >>>>>>> For me, the more core of an export it is, the stronger the reason it
> > >>>>>>> should be GPL. FWIW, I don't think exporting task_work functionality is
> > >>>>
> > >>>>>>
> > >>>>>> Yeah, I'm not too fond of that part as well. I don't think we want to
> > >>>>>> give modules the ability to mess with task work. This is just asking for
> > >>>>>> trouble.
> > >>>>>>
> > >>>>>
> > >>>>> Ok, maybe we need to reframe the problem then.
> > >>>>>
> > >>>>> Currently fput(), and hence filp_close(), take control away from kernel
> > >>>>> threads in that they cannot be sure that a "close" has actually
> > >>>>> completed.
> > >>>>>
> > >>>>> This is already a problem for nfsd. When renaming a file, nfsd needs to
> > >>>>> ensure any cached "open" that it has on the file is closed (else when
> > >>>>> re-exporting an NFS filesystem it can result in a silly-rename).
> > >>>>>
> > >>>>> nfsd currently handles this case by calling flush_delayed_fput(). I
> > >>>>> suspect you are no more happy about exporting that than you are about
> > >>>>> exporting task_work_run(), but this solution isn't actually 100%
> > >>>>> reliable. If some other thread calls flush_delayed_fput() between nfsd
> > >>>>> calling filp_close() and that same nfsd calling flush_delayed_fput(),
> > >>>>> then the second flush can return before the first flush (in the other
> > >>>>> thread) completes all the work it took on.
> > >>>>>
> > >>>>> What we really need - both for handling renames and for avoiding
> > >>>>> possible memory exhaustion - is for nfsd to be able to reliably wait for
> > >>>>> any fput() that it initiated to complete.
> > >>>>>
> > >>>>> How would you like the VFS to provide that service?
> > >>>>
> > >>>> Since task_work happens in the context of your task already, why not
> > >>>> just have a way to get it stashed into a list when final fput is done?
> > >>>> This avoids all of this "let's expose task_work" and using the task list
> > >>>> for that, which seems kind of pointless as you're just going to run it
> > >>>> later on manually anyway.
> > >>>>
> > >>>> In semi pseudo code:
> > >>>>
> > >>>> bool fput_put_ref(struct file *file)
> > >>>> {
> > >>>> return atomic_dec_and_test(&file->f_count);
> > >>>> }
> > >>>>
> > >>>> void fput(struct file *file)
> > >>>> {
> > >>>> if (fput_put_ref(file)) {
> > >>>> ...
> > >>>> }
> > >>>> }
> > >>>>
> > >>>> and then your nfsd_file_free() could do:
> > >>>>
> > >>>> ret = filp_flush(file, id);
> > >>>> if (fput_put_ref(file))
> > >>>> llist_add(&file->f_llist, &l->to_free_llist);
> > >>>>
> > >>>> or something like that, where l->to_free_llist is where ever you'd
> > >>>> otherwise punt the actual freeing to.
> > >>>
> > >>> Should probably have the put_ref or whatever helper also init the
> > >>> task_work, and then reuse the list in the callback_head there. Then
> > >>> whoever flushes it has to call ->func() and avoid exposing ____fput() to
> > >>> random users. But you get the idea.
> > >>
> > >> Interesting ideas - thanks.
> > >>
> > >> So maybe the new API would be
> > >>
> > >> fput_queued(struct file *f, struct llist_head *q)
> > >> and
> > >> flush_fput_queue(struct llist_head *q)
> > >>
> > >> with the meaning being that fput_queued() is just like fput() except
> > >> that any file needing __fput() is added to the 'q'; and that
> > >> flush_fput_queue() calls __fput() on any files in 'q'.
> > >>
> > >> So to close a file nfsd would:
> > >>
> > >> fget(f);
> > >> flip_close(f);
> > >> fput_queued(f, &my_queue);
> > >>
> > >> though possibly we could have a
> > >> filp_close_queued(f, q)
> > >> as well.
> > >>
> > >> I'll try that out - but am happy to hear alternate suggestions for names :-)
> > >>
> > >
> > > Actually .... I'm beginning to wonder if we should just use
> > > __fput_sync() in nfsd.
> > > It has a big warning about not doing that blindly, but the detail in the
> > > warning doesn't seem to apply to nfsd...
> >
> > If you can do it from the context where you do the filp_close() right
> > now, then yeah there's no reason to over-complicate this at all... FWIW,
>
> As long as nfsd doesn't care that it may get stuck on umount or
> ->release...
I think we do *care* about getting stuck. But I don't think we would
*expect* to get stuck..
I had a look at varous ->release function. Quite few do fsync or
similar which isn't a problem. nfsd often waits for writes to complete.
Some lock the inode, which again is something that nfsd threads often
do.
Is there something special that ->release might do but that other
filesystem operation don't do?
I'd really like to understand why __fput is so special that we often
queue it to a separate thread.
>
> > the reason task_work exists is just to ensure a clean context to perform
> > these operations from the task itself. The more I think about it, it
> > doesn't make a lot of sense to utilize it for this purpose, which is
> > where my alternate suggestion came from. But if you can just call it
> > directly, then that makes everything much easier.
>
> And for better or worse we already expose __fput_sync(). We've recently
> switched close(2) over to it as well as it was needlessly punting to
> task work.
>
exit_files() would be another good candidate for using __fput_sync().
Oleg Nesterov has reported problems when a process which a large number
of files exits - this currently puts lots of entries on the task_works
lists. If task_work_cancel is then called before those are all dealt
with, it can have a long list to search while holding a hot lock. (I
hope I got that description right).
Thanks,
NeilBrown
