RE: [PATCH] ceph: fix hanging __ceph_get_caps() with stale `mds_wanted`

[Viacheslav Dubeyko <Slava.Dubeyko@ibm.com>]

On Wed, 2026-04-29 at 06:27 +0200, Max Kellermann wrote:
> On Tue, Apr 28, 2026 at 8:46 PM Viacheslav Dubeyko
> <Slava.Dubeyko@ibm.com> wrote:
> > >                       flags |= NON_BLOCKING;
> > >                       while (!(ret = try_get_cap_refs(inode, need, want,
> > >                                                       endoff, flags, &_got))) {
> > > +                             static const unsigned long wait_timeout = 5 * HZ;
> > 
> > Why exactly 5 * HZ? What is the basis for this timeout? Could we re-use any
> > available timeouts in CephFS declarations?
> 
> It is an arbitrary timeout, long enough to avoid unnecessary wakeups
> in regular situations where we're really waiting for a capability, but
> short enough to avoid disrupting the service. If you prefer another
> number, say it, and I'll change it.
> 
> On our servers, this hang bug occurs very rarely. Sometimes, weeks go
> by without a hang, and sometimes twice a day, but no more. When it
> happens, I thought it's fine to wait 5 seconds before attempting to
> recover.
> 
> Note that this is still a minimal workaround, not a proper fix, as I
> wrote. The proper fix would be much more intrusive (and of course go
> without any hard-coded timeouts and arbitrary wakeups).

I have nothing in mind as another value. I think it will be good to introduce
the declaration of timeout:

#define CEPH_GET_CAPS_WAIT_TIMEOUT (5 * HZ)

And, probably, we need to introduce it with other timeouts [1].

> 
> > >                       if (ret == -EUCLEAN) {
> > > -                             /* session was killed, try renew caps */
> > > -                             ret = ceph_renew_caps(inode, flags);
> > > +                             /* session was killed or a waited cap
> > > +                              * request needs a retry */
> > > +                             ret = ceph_renew_caps(inode, flags & CEPH_FILE_MODE_MASK);
> > 
> > Frankly speaking, I don't quite follow why do we need to add flags &
> > CEPH_FILE_MODE_MASK?
> 
> Oh, this is an unrelated fix.  This "flags" variable contains internal
> flags that are not understood by ceph_renew_caps() (i.e.
> CHECK_FILELOCK and NON_BLOCKING) and shouldn't really be passed there.
> 
> I'll remove that change from the patch for v2 because while I believe
> it's correct, it has nothing to do with this bug.  I'll post v2 when
> we agree on the rest of the patch.

I assume that flags & CEPH_FILE_MODE_MASK cleans the NON_BLOCKING flag. Am I
correct? So, it looks like this fix also necessary. Am I right?

The patch make sense to me.

Thanks,
Slava.

[1]
https://elixir.bootlin.com/linux/v7.0.1/source/include/linux/ceph/libceph.h#L72