Re: [PATCH 1/3] direct-io: only inc/dec inode->i_dio_count for file systems

[Jens Axboe]

On 04/15/2015 04:36 PM, Dave Chinner wrote:
> On Wed, Apr 15, 2015 at 04:01:36PM -0600, Jens Axboe wrote:
> > do_blockdev_direct_IO() increments and decrements the inode
> > ->i_dio_count for each IO operation. It does this to protect against
> > truncate of a file. Block devices don't need this sort of protection.
> >
> > For a capable multiqueue setup, this atomic int is the only shared
> > state between applications accessing the device for O_DIRECT, and it
> > presents a scaling wall for that. In my testing, as much as 30% of
> > system time is spent incrementing and decrementing this value. A mixed
> > read/write workload improved from ~2.5M IOPS to ~9.6M IOPS, with
> > better latencies too. Before:
> .....
> > diff --git a/fs/inode.c b/fs/inode.c
> > index f00b16f45507..c4901c40ad65 100644
> > --- a/fs/inode.c
> > +++ b/fs/inode.c
> > @@ -1946,18 +1946,31 @@ void inode_dio_wait(struct inode *inode)
> >   EXPORT_SYMBOL(inode_dio_wait);
> >
> >   /*
> > - * inode_dio_done - signal finish of a direct I/O requests
> > + * inode_dio_begin - signal start of a direct I/O requests
> >    * @inode: inode the direct I/O happens on
> >    *
> >    * This is called once we've finished processing a direct I/O request,
> >    * and is used to wake up callers waiting for direct I/O to be quiesced.
> >    */
> > -void inode_dio_done(struct inode *inode)
> > +void inode_dio_inc(struct inode *inode)
>
> function name does not match docbook comment....

Oops, will fix that up.

> > +{
> > +	atomic_inc(&inode->i_dio_count);
> > +}
> > +EXPORT_SYMBOL(inode_dio_inc);
> > +
> > +/*
> > + * inode_dio_dec - signal finish of a direct I/O requests
> > + * @inode: inode the direct I/O happens on
> > + *
> > + * This is called once we've finished processing a direct I/O request,
> > + * and is used to wake up callers waiting for direct I/O to be quiesced.
> > + */
> > +void inode_dio_dec(struct inode *inode)
> >   {
> >   	if (atomic_dec_and_test(&inode->i_dio_count))
> >   		wake_up_bit(&inode->i_state, __I_DIO_WAKEUP);
> >   }
> > -EXPORT_SYMBOL(inode_dio_done);
> > +EXPORT_SYMBOL(inode_dio_dec);
>
> Bikeshedding: I think this would be better suited to inode_dio_begin()
> and inode_dio_end() because now we are trying to say "this is where
> the DIO starts, and this is where it ends". It's not really
> "reference counting" interface, we're trying to annotate the
> boundaries of where DIO iis protected against truncate....

I don't really care, if people like begin/end more than inc/dec, I'm 
happy with that.

> And, realistically, if we are pushing this up into the filesystems
> again, we should push it up into *all* filesystems and get rid of it
> completely from the DIO layer. That way no new twisty passages in
> the direct IO code are needed.

Lets please keep that for a potential round 2. It's not like I'm piling 
lots of hacks on, it's two one-liner changes. It's not adding a lot to 
the entropy of direct-io.c. I've been carrying this patch for years now, 
I really don't want to sign up for futzing around in direct-io.c, nor is 
that a reasonable requirement imho.

--
Jens Axboe

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/