On 22 Mar, Konstantin Belousov wrote:
> On Sun, Mar 22, 2015 at 02:37:09PM +0100, Jilles Tjoelker wrote:
>> On Sat, Mar 21, 2015 at 08:49:00PM +1100, Bruce Evans wrote:
>> > On Sat, 21 Mar 2015, Xin LI wrote:
>>
>> > > Log:
>> > > Disable timestamping on devfs read/write operations by default.
>>
>> > > Currently we update timestamps unconditionally when doing read or
>> > > write operations. This may slow things down on hardware where
>> > > reading timestamps is expensive (e.g. HPET, because of the default
>> > > vfs.timestamp_precision setting is nanosecond now) with limited
>> > > benefit.
>>
>> > > A new sysctl variable, vfs.devfs.dotimes is added, which can be
>> > > set to non-zero value when the old behavior is desirable.
>>
>> > I don't like this. It defaults to non-POSIX-conformant behaviour...
>>
>> > The slowness is mostly from no delayed update of times in devfs.
>> > Switching vfs.timestamp_precision to a hardware timecounter would
>> > have been even more expensive for regular files if file systems
>> > didn't have delayed updates. The assumption that vfs_timestamp()
>> > doesn't use a slow timecounter was so often satisfied that no one
>> > missed devfs also not supporting mounting with -noatime.
>>
>> > Delayed updates are even easier to implement for devfs than for disk
>> > file systems the times never need to be written to disk. A slow update
>> > is still wasteful for atimes, but not as bad as for disk file systems
>> > since it doesn't trigger a slower sync to disk.
>>
>> Yes, I think implementing delayed updates is the right solution to this
>> problem. This way, only stat and last close will need to read the clock.
>> No configuration option will be needed.
>>
>> A subtle difference with most other file systems is that devfs nodes
>> often stay open for very long, so the timestamps will usually come from
>> stat() calls, which may be much later than the actual read or write.
>> Still that is better than not updating timestamps at all.
>>
>> > > ...
>> > > Modified: head/sys/fs/devfs/devfs_vnops.c
>> > > ==============================================================================
>> > > --- head/sys/fs/devfs/devfs_vnops.c Sat Mar 21 00:21:30 2015
>> > > (r280307)
>> > > +++ head/sys/fs/devfs/devfs_vnops.c Sat Mar 21 01:14:11 2015
>> > > (r280308)
>> > > ...
>> > > @@ -1700,7 +1708,8 @@ devfs_write_f(struct file *fp, struct ui
>> > > resid = uio->uio_resid;
>> > >
>> > > error = dsw->d_write(dev, uio, ioflag);
>> > > - if (uio->uio_resid != resid || (error == 0 && resid != 0)) {
>> > > + if (devfs_dotimes &&
>> > > + (uio->uio_resid != resid || (error == 0 && resid != 0))) {
>> > > vfs_timestamp(&dev->si_ctime);
>> > > dev->si_mtime = dev->si_ctime;
>> > > }
>>
>> > An old bug is evident in the diff. Writing shouldn't change the ctime.
>>
>> That is not a bug. POSIX unambiguously requires write() to update both
>> mtime and ctime.
> Does POSIX ever say anything about special files ?
>
> Devfs already has non-POSIX behaviour, e.g. write on one mount is
> reflected as mtime/ctime update on all mounts. On reboot, the time
> stamps are re-created, i.e. changes are not persistent.
>
> I think the deviations may be summarrized as 'devfs mtime is useless for
> the usual mtime purposes'. From this PoV, I have no objections to the
> patch. Doing extra work with delayed updates of times, which might
> be somewhat non-trivial, is a feature with non-obvious gain.
It's not totally worthless. I think the mtime on tty devices is used to
calculate the idle time that is printed by the w command. We just don't
need nanosecond accuracy for that.
_______________________________________________
svn-src-head@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/svn-src-head
To unsubscribe, send any mail to "svn-src-head-unsubscr...@freebsd.org"
