Date: Thu, 18 Aug 2016 23:39:42 +0200 From: Jilles Tjoelker <jilles@stack.nl> To: Konstantin Belousov <kostikbel@gmail.com> Cc: Konstantin Belousov <kib@FreeBSD.org>, src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: Re: svn commit: r304176 - in head: include lib/libc/sys sys/compat/freebsd32 sys/kern Message-ID: <20160818213941.GA1539@stack.nl> In-Reply-To: <20160818075316.GW83214@kib.kiev.ua> References: <201608151908.u7FJ8phh091939@repo.freebsd.org> <20160817202040.GA21263@stack.nl> <20160818075316.GW83214@kib.kiev.ua>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, Aug 18, 2016 at 10:53:16AM +0300, Konstantin Belousov wrote:
> On Wed, Aug 17, 2016 at 10:20:40PM +0200, Jilles Tjoelker wrote:
> > On Mon, Aug 15, 2016 at 07:08:51PM +0000, Konstantin Belousov wrote:
> > > /* ISO/IEC 9945-1: 1996 */
> > > #if __POSIX_VISIBLE >= 199506 || __XSI_VISIBLE
> > > int fsync(int);
> > > +int fdatasync(int);
> > >
> > > /*
> > > * ftruncate() was in the POSIX Realtime Extension (it's used for shared
> > Apparently these functions were added closely enough in time that they
> > can stay together here :)
> Is this a form of suggestion to use other value for POSIX_VISIBLE ?
No, no change is necessary.
> > > +#if 0
> > > + if (!fullsync)
> > > + /* XXXKIB: compete outstanding aio writes */;
> > Under the _POSIX_SYNCHRONIZED_IO option, completing outstanding I/O
> > requests is in fact required for fsync() as well. The fdatasync()
> > function is completely under the _POSIX_SYNCHRONIZED_IO option.
> > We do not implement this option, but keeping fdatasync()'s guarantees a
> > subset of fsync()'s guarantees seems sensible.
> I will consider this if and when the AIO flush would be implemented.
> I looked at the AIO code to estimate the needed work, but did not started
> coding and quite possible it would be postponed.
OK.
> > > +%% fdatasync vp L L L
> > > +
> > > +vop_fdatasync {
> > > + IN struct vnode *vp;
> > > + IN struct thread *td;
> > > +};
> > A waitfor parameter like in vop_fsync may be useful to implement
> > aio_fsync(O_DSYNC) later on.
> I really do not see how would it be. aio_fsync(O_DSYNC) is equivalent
> to fdatasync(2) in the async context. Completion of the aio request
> indicates that virtual fdatasync(2) execution did finished in that
> context. The use of waitfor in the VOP_FSYNC() is to allow syncer
> to initiate flush without waiting, or getting notification for the
> completion.
> Could you, please, elaborate ? If the KPI change is needed there, it
> is obviously desirable to make it right before MFC to stable.
Oh, you're right. No change is needed. A MNT_NOWAIT fsync operation does
not provide any feedback when it is complete, so it is not useful for
AIO. In fact, aio_fsync() eventually calls VOP_FSYNC with MNT_WAIT on a
worker thread.
--
Jilles Tjoelker
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20160818213941.GA1539>