On Thu, May 25, 2006 at 07:27:13PM +1000, Nick Piggin wrote:
> Wu Fengguang wrote:
>
> >
> >>Ah, I'm just having a little dig ;) (as you do with me when I break the
> >>radix-tree). No, it's great to see you're picking up the readahead stuff,
> >>however I guess one has to be fairly careful with the radix-tree though...
> >>
> >
> >I had a feeling that there were too few reviews, and with this
> >feeling, have been improving the code/doc/patch for easier reviews.
> >
> >Thank you very much for the comments, most of them are really valuable
> >ones. I'll answer them one by one and try to solve the problems ASAP.
> >
>
> OK: just make sure to cc lkml (I forgot to with this thread)...
ok.
> >
> >>I wouldn't like to see you drop it all, however if Wu can come up with a
> >>patchset minus radix tree changes I think it would have a better chance.
> >>
> >
> >I feel very sorry for the collision with your patches.
> >I have been worried about it for some time.
> >Maybe it's time for us to address it in some constructive way?
> >
>
> Well I think the direct data patches are fairly important, and are smallish
> and standalone. Ie. they'll most likely get in before any readahead stuff,
> so you'll have to look at porting patches to them.
>
> I can help with that if/when the time comes. However, as I've said, I would
> drop them for the moment and just focus on the core readahead patchset.
Oh please keep them, I have fixed the code to work with the direct
data :)
Now I have removed the radix tree look-aside cache, and cut down the
new radix tree functions to three core ones:
- __radix_tree_lookup_parent()
- radix_tree_scan_hole_backward()
- radix_tree_scan_hole()
In the hope that it be easier for you to work with.
> >>Wu: if someone does report that it causes poor CPU efficiency, that won't
> >>be cause for the patchset to be dropped: on the contrary it might provide
> >>some good evidence that the radix-tree work is needed. And then we can
> >>look at as a problem by itself.
> >>
> >
> >You are guessing right: I've been nervous about the CPU efficiency of
> >the context based method since the early stage of the patch. OMG it
> >would be great if I get this advice early ;)
> >
>
> I guess this was the point I had been trying to get at. Maybe I wasn't
> explicit :\
;-)
Wu