Re: [PATCH] Changed blk trace msgs to directly use relay buffer

[Jens Axboe <jens.axboe@oracle.com>]

On Wed, May 28 2008, Alan D. Brunelle wrote:
> Jens Axboe wrote:
> > On Wed, May 28 2008, Alan D. Brunelle wrote:
> >> Jens Axboe wrote:
> >>> On Tue, May 27 2008, Alan D. Brunelle wrote:
> >>>
> >>>> From 43c8ea2b78f31d7ccd349384a9a2084e787aafc1 Mon Sep 17 00:00:00 2001
> >>>> From: Alan D. Brunelle <alan.brunelle@hp.com>
> >>>> Date: Tue, 27 May 2008 10:32:36 -0400
> >>>> Subject: [PATCH] Changed blk trace msgs to directly use relay buffer
> >>>>
> >>>> Allows for SMP-usage without corruption, and removes an extra copy at
> >>>> the expense of copying extra bytes. Reduced message size from 1024 to 128.
> >>> Or, alternatively, something like the below. Then we don't
> >>> unconditionally reserve and copy 128 bytes for each message, at the
> >>> cost 128 bytes per-cpu per trace.
> >> I looked into something like this, but thought the added complexity
> >> wasn't worth it. Besides the extra per-cpu stuff, you also have an
> >> extra memcopy involved - in my patch you print directly into the relay
> >> buffer.  I figure that /if/ copying (128-msg_size) extra bytes is too
> >> much, one could always shrink the 128 down further. [I would think 64
> >> bytes is probably ok.]
> >>
> >> I'd bet that the reduced complexity, and skipping the extra memcopy
> >> more than offsets having to copy a few extra bytes...
> > 
> > The complexity is the same imho, both versions are fairly trivial.
> > I wasn't out to optimize this in a memory copy sense. To me the most
> > precious resource is the data stream to the app, and 128 bytes
> > is probably 6 times larger than the normal message would be. With
> > the actual trace structure, we are down to about 3 times the byte
> > size.
> > 
> > So it was just an idea, I don't care much either way. With 128 bytes,
> > we could just put the buffer on the stack (and still do the copy to
> > the relay buffer). The per-cpu buffers has the advantage that we
> > could grow the size easily if we wanted to.
> > 
> > So, given everything, which do you prefer?
> > 
> 
> Given your prioritizing of relay-copying over kernel-copying, I think
> you're solution is better (and more flexible going forward).

OK good, it's committed and going upstream soon(ish).

-- 
Jens Axboe