Re: md question re: max_hw_sectors_kb

[NeilBrown <neilb@suse.de>]

On Fri, 6 May 2011 14:24:34 +1000 NeilBrown <neilb@suse.de> wrote:

> On Tue, 03 May 2011 15:20:46 -0500 Michael Reed <mdr@sgi.com> wrote:
> 
> > Resending to linux-raid.
> > 
> > 
> > Hi Neil,
> > 
> > My name is Mike Reed.  I work at SGI.  I've been looking at a performance
> > issue with md writes and have tracked it down to the md device not having
> > a high enough initial max_hw_sectors_kb setting.
> > 
> > There is code in blk_queue_make_request() which lowers the default value
> > from INT_MAX to BLK_SAFE_MAX_SECTORS, which is 255.  This is generally
> > lower than all the underlying devices with which I use md.
> > 
> > As md appears to be a stacking driver, i.e., it calls disk_stack_limits()
> > for each member of a volume, it would seem reasonable for md to use the,
> > INT_MAX setting for max_hw_sectors_kb instead of BLK_SAFE_MAX_SECTORS.
> > 
> > I have tried this, and have observed that md correctly limits the md device's
> > max_sectors_kb to the value of the underlying devices in my mirror volume.
> > 
> > Is this the correct way to address this issue?
> > 
> > Applies to linux-2.6.39-rc4.
> > 
> > Signed-off-by: Michael Reed <mdr@sgi.com>
> > 
> > 
> > --- drivers/md/md.c	2011-04-18 21:26:00.000000000 -0700
> > +++ drivers/md/md.c.new	2011-04-21 15:53:11.452536201 -0700
> > @@ -4328,6 +4328,7 @@ static int md_alloc(dev_t dev, char *nam
> >  	mddev->queue->queuedata = mddev;
> >  
> >  	blk_queue_make_request(mddev->queue, md_make_request);
> > +	blk_queue_max_sectors(mddev->queue, INT_MAX);
> >  
> >  	disk = alloc_disk(1 << shift);
> >  	if (!disk) {
> > 
> 
> Hi Mike,
>  sorry for not responding to this earlier - it seemed to keep falling through
>  cracks for some reason :-(
> 
> Yes, I completely agree with your analysis and think you patch is correct and
> useful.
> I will queue it up for the next merge window.
> 
> thanks,
> NeilBrown

...except of course that it should be
     blk_queue_max_hw_sectors(......)
                   ^^

because blk_queue_max_sectors was renamed back in 2.6.34 :-)

NeilBrown