Re: [PATCH 3/7] IB/srpt: Change default behavior from using SRQ to not using SRQ

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Tue, 2017-10-10 at 18:44 +0300, Leon Romanovsky wrote:
> On Tue, Oct 10, 2017 at 11:13:22AM -0400, Doug Ledford wrote:
> > On Tue, 2017-10-10 at 18:00 +0300, Leon Romanovsky wrote:
> > > On Tue, Oct 10, 2017 at 10:34:19AM -0400, Doug Ledford wrote:
> > > > On 10/10/2017 12:14 AM, Leon Romanovsky wrote:
> > > > > 
> > > 
> > > Doug,
> > > But my question was "when" and not "how". When should I set this
> > > parameter to true?
> > 
> > Whenever you want.  When do you set the maximum TCP window size to
> > 4MB?
> >  If you have an issue with things being one way, you try another.
> > Things like this are system specific and often there is no
> > universal
> > answer, especially since the answer here could very easily depend
> > on
> > things like how many targets you are connecting to, how fast those
> > targets are, etc.
> 
> Bart clearly mentioned disadvantages of XRQ and left me wonder why
> user
> needs to enable it anyway. This is what I'm asking and this is what
> I'm
> hoping to see in the commit message.

That's more than a commit message needs to be IMO.

> I know little about SRP and by asking the questions, I'm trying to
> fill
> the gaps.

It's not really SRP related.  As Bart points out in his next message,
it's a tradeoff between lower resource consumption and higher
contention.  It's a general tradeoff that is common to RDMA.

> > 
> > > > 
> > > > > In the commit message, you mentioned disadvantages of using
> > > > > SRQ
> > > > > is a
> > > > > default and among them - locks contention, which can be
> > > > > changed
> > > > > in the
> > > > > future. Won't it mean that users stuck with current default,
> > > > > because
> > > > > change of default will "break" their scripts?
> > > > 
> > > > No, it won't.  If you change the default, you don't remove the
> > > > variable,
> > > > you just change what its setting is.  Then existing modprobe.d
> > > > files
> > > > become redundant, but nothing breaks.  People that don't want
> > > > the
> > > > new
> > > > setting add a new file to the modprobe.d directory to change
> > > > the
> > > > option.
> > > 
> > > Not accurate, now I won't set any parameter because I'm relying
> > > on
> > > the
> > > fact that the default is without SRQ. Once the default will be
> > > changed,
> > > it will break my assumption.
> > 
> > So?  Your assumptions are never a reason to prevent ongoing
> > development.  By this argument, we would never have been able to
> > turn
> > on SCSI Multi-queue by default because when it was first added to
> > the
> > kernel it was off by default and people might have made an
> > assumption
> > that we would later break when we turned it on by default.
> 
> I'm not fully understand the example, The transition to MQ was long
> process and one of the difficulties was requirement to smoothly
> transit
> all devices in such way that users won't be affected.

The difficulty was making sure it was robust.

>  This is why it was
> disabled by default.

It was off by default because they wanted it to get testing and bake
time before enabling it by default.  However, the reason they did it is
really immaterial to my point: we can change defaults and that is not
considered "breaking user's assumptions".  The only way we would break
user's system in this specific situation is if we removed the module
option.  That would break working setups that had a modprobe.d file to
set the variable.  Changing what the variable is by default doesn't do
that.

> Thanks
> 
> > 
> > --
> > Doug Ledford <dledford@xxxxxxxxxx>
> >     GPG KeyID: B826A3330E572FDD
> >     Key fingerprint = AE6B 1BDA 122B 23B4 265B  1274 B826 A333 0E57
> > 2FDD
> > 
-- 
Doug Ledford <dledford@xxxxxxxxxx>
    GPG KeyID: B826A3330E572FDD
    Key fingerprint = AE6B 1BDA 122B 23B4 265B  1274 B826 A333 0E57 2FDD

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



[Index of Archives]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Photo]     [Yosemite News]     [Yosemite Photos]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux