Re: [PATCH RFC 1/1] nfsd: Initial implementation of NFSv4 Courteous Server

["J. Bruce Fields" <bfields@fieldses.org>]

On Wed, Jun 30, 2021 at 12:13:35PM -0700, dai.ngo@oracle.com wrote:
> 
> On 6/30/21 11:55 AM, J. Bruce Fields wrote:
> >On Wed, Jun 30, 2021 at 11:49:18AM -0700, dai.ngo@oracle.com wrote:
> >>On 6/30/21 11:05 AM, J. Bruce Fields wrote:
> >>>On Wed, Jun 30, 2021 at 10:51:27AM -0700, dai.ngo@oracle.com wrote:
> >>>>>On 6/28/21 1:23 PM, J. Bruce Fields wrote:
> >>>>>>where ->fl_expire_lock is a new lock callback with second
> >>>>>>argument "check"
> >>>>>>where:
> >>>>>>
> >>>>>>     check = 1 means: just check whether this lock could be freed
> >>>>Why do we need this, is there a use case for it? can we just always try
> >>>>to expire the lock and return success/fail?
> >>>We can't expire the client while holding the flc_lock.  And once we drop
> >>>that lock we need to restart the loop.  Clearly we can't do that every
> >>>time.
> >>>
> >>>(So, my code was wrong, it should have been:
> >>>
> >>>
> >>>	if (fl->fl_lops->fl_expire_lock(fl, 1)) {
> >>>		spin_unlock(&ct->flc_lock);
> >>>		fl->fl_lops->fl_expire_locks(fl, 0);
> >>>		goto retry;
> >>>	}
> >>>
> >>>)
> >>This is what I currently have:
> >>
> >>retry:
> >>                 list_for_each_entry(fl, &ctx->flc_posix, fl_list) {
> >>                         if (!posix_locks_conflict(request, fl))
> >>                                 continue;
> >>
> >>                         if (fl->fl_lmops && fl->fl_lmops->lm_expire_lock) {
> >>                                 spin_unlock(&ctx->flc_lock);
> >>                                 ret = fl->fl_lmops->lm_expire_lock(fl, 0);
> >>                                 spin_lock(&ctx->flc_lock);
> >>                                 if (ret)
> >>                                         goto retry;
> >We have to retry regardless of the return value.  Once we've dropped
> >flc_lock, it's not safe to continue trying to iterate through the list.
> 
> Yes, thanks!
> 
> >
> >>                         }
> >>
> >>                         if (conflock)
> >>                                 locks_copy_conflock(conflock, fl);
> >>
> >>>But the 1 and 0 cases are starting to look pretty different; maybe they
> >>>should be two different callbacks.
> >>why the case of 1 (test only) is needed,  who would use this call?
> >We need to avoid dropping the spinlock in the case there are no clients
> >to expire, otherwise we'll make no forward progress.
> 
> I think we can remember the last checked file_lock and skip it:

I doubt that works in the case there are multiple locks with
lm_expire_lock set.

If you really don't want another callback here, maybe you could set some
kind of flag on the lock.

At the time a client expires, you're going to have to walk all of its
locks to see if anyone's waiting for them.  At the same time maybe you
could set an FL_EXPIRABLE flag on all those locks, and test for that
here.

If the network partition heals and the client comes back, you'd have to
remember to clear that flag again.

--b.

> retry:
>                 list_for_each_entry(fl, &ctx->flc_posix, fl_list) {
>                         if (!posix_locks_conflict(request, fl))
>                                 continue;
> 
>                         if (checked_fl != fl && fl->fl_lmops &&
>                                         fl->fl_lmops->lm_expire_lock) {
>                                 checked_fl = fl;
>                                 spin_unlock(&ctx->flc_lock);
>                                 fl->fl_lmops->lm_expire_lock(fl);
>                                 spin_lock(&ctx->flc_lock);
>                                 goto retry;
>                         }
> 
>                         if (conflock)
>                                 locks_copy_conflock(conflock, fl);
> 
> -Dai
> 
> >
> >--b.