On Fri, Feb 28, 2014 at 07:01:50AM +1100, Dave Chinner wrote:
> On Thu, Feb 27, 2014 at 09:08:46AM -0500, Brian Foster wrote:
> > On Thu, Feb 27, 2014 at 08:51:14PM +1100, Dave Chinner wrote:
> > > From: Dave Chinner <dchinner@xxxxxxxxxx>
> > >
> > Hmm, I replied to this one in the previous thread, but now I notice that
> > it apparently never made it to the list. Dave, did you happen to see
> > that in your inbox? Anyways, I had a couple minor comments/questions
> > that I'll duplicate here (which probably don't require another
> > repost)...
> No, I didn't.
> [snip typos that need fixing]
> > > diff --git a/repair/prefetch.c b/repair/prefetch.c
> > > index aee6342..7d3efde 100644
> > > --- a/repair/prefetch.c
> > > +++ b/repair/prefetch.c
> > > @@ -866,6 +866,48 @@ start_inode_prefetch(
> > > return args;
> > > }
> > >
> > A brief comment before the prefetch_ag_range bits that explain the
> > implicit design constraints (e.g., throttle prefetch based on
> > processing) would be nice. :)
> Can do.
* prefetch_ag_range runs a prefetch-and-process loop across a range of AGs. It
* begins with @start+ag, and finishes with @end_ag - 1 (i.e. does not prefetch
* or process @end_ag). The function starts prefetch on the first AG, then loops
* starting prefetch on the next AG and then blocks processing the current AG as
* the prefetch queue brings inodes into the processing queue.
* There is only one prefetch taking place at a time, so the prefetch on the
* next AG only starts once the current AG has been completely prefetched. Hence
* the prefetch of the next AG will start some time before the processing of the
* current AG finishes, ensuring that when we iterate an start processing the
* next AG there is already a significant queue of inodes to process.
* Prefetch is done this way to prevent it from running too far ahead of the
* processing. Allowing it to do so can cause cache thrashing, where new
* prefetch causes previously prefetched buffers to be reclaimed before the
* processing thread uses them. This results in reading all the inodes and
* metadata twice per phase and it greatly slows down the processing. Hence we
* have to carefully control how far ahead we prefetch...