Re: [PATCH 05/14] xfs: refactor quota expiration timer modification

From: "Darrick J. Wong" <darrick.wong@oracle.com>
To: Eric Sandeen <sandeen@sandeen.net>
Cc: linux-xfs@vger.kernel.org
Subject: Re: [PATCH 05/14] xfs: refactor quota expiration timer modification
Date: Wed, 12 Feb 2020 21:33:18 -0800	[thread overview]
Message-ID: <20200213053318.GR6874@magnolia> (raw)
In-Reply-To: <e8d6702c-9e39-6e5e-e910-739b9b4558a3@sandeen.net>

On Wed, Feb 12, 2020 at 09:32:33PM -0600, Eric Sandeen wrote:
> 
> 
> On 2/12/20 9:27 PM, Eric Sandeen wrote:
> > On 2/12/20 7:46 PM, Darrick J. Wong wrote:
> >> On Wed, Feb 12, 2020 at 05:57:24PM -0600, Eric Sandeen wrote:
> >>> On 12/31/19 7:11 PM, Darrick J. Wong wrote:
> >>>> From: Darrick J. Wong <darrick.wong@oracle.com>
> > 
> > ...
> > 
> >>>>  	} else {
> >>>>  		if (!over) {
> >>>> -			d->d_btimer = 0;
> >>>> +			xfs_quota_clear_timer(&d->d_btimer);
> >>>
> >>> yeah that's a very fancy way to say "= 0" ;)
> >>>
> >>> ...
> >>
> >> Yes, that's a fancy way to assign zero.  However, consider that for
> >> bigtime support, I had to add an incore quota timer so that timers more
> >> or less fire when they're supposed to, and now there's a function to
> >> convert the incore timespec64 value to whatever is ondisk:
> >>
> >> /* Clear a quota grace period expiration timer. */
> >> static inline void
> >> xfs_quota_clear_timer(
> >> 	struct xfs_disk_dquot	*ddq,
> >> 	time64_t		*itimer,
> >> 	__be32			*dtimer)
> >> {
> >> 	struct timespec64	tv = { 0 };
> >>
> >> 	*itimer = tv.tv_sec;
> >> 	xfs_dquot_to_disk_timestamp(ddq, dtimer, &tv);
> >> }
> >>
> >> It was at *that* point in the patchset that it seemed easier to call a
> >> small function three times than to open-code this three times.
> > 
> > +void
> > +xfs_dquot_to_disk_timestamp(
> > +	__be32			*dtimer,
> > +	const struct timespec64	*tv)
> > +{
> > +	*dtimer = cpu_to_be32(tv->tv_sec);
> > +}
> > 
> >  static inline void
> >  xfs_quota_clear_timer(
> > +	time64_t		*itimer,
> >  	__be32			*dtimer)
> >  {
> > -	*dtimer = cpu_to_be32(0);
> > +	struct timespec64	tv = { 0 };
> > +
> > +	*itimer = tv.tv_sec;
> > +	xfs_dquot_to_disk_timestamp(dtimer, &tv);
> >  } 
> > 
> > xfs_quota_clear_timer(&dqp->q_btimer, &d->d_btimer);
> > 
> > That's still a very fancy way of saying:
> > 
> >         dqp->q_btimer = d->d_btimer = 0;
> > 
> > I think?  Can't really see what value the timespec64 adds here.
> > 
> > -Eric
> > 
> 
> Actually,
> 
>  xfs_quota_set_timer(
> +	time64_t		*itimer,
>  	__be32			*dtimer,
>  	time_t			limit)
>  {
> -	time64_t		new_timeout;
> +	struct timespec64	tv = { 0 };
>  
> -	new_timeout = xfs_dquot_clamp_timer(get_seconds() + limit);
> -	*dtimer = cpu_to_be32(new_timeout);
> +	tv.tv_sec = xfs_dquot_clamp_timer(ktime_get_real_seconds() + limit);
> +	*itimer = tv.tv_sec;
> +	xfs_dquot_to_disk_timestamp(dtimer, &tv);
>  }
> 
> I'm not sure why there's a timespec64 here either.  Isn't everything
> we're dealing with on timers in seconds, using only tv_sec, and time64_t would
> suffice instead of using a timespec64 just to carry around a seconds value?

Yes, the grace periods recorded in the timer fields in dquot 0 are
intervals measured in seconds.

However, for dquot != 0, the timer fields store the time of the
expiration, so I settled on timespec64 as the incore structure so that
XFS consistently uses struct timespec64 to represent specific points in
time.

(That and time64_t doesn't exist in userspace.)

--D

> -Eric