Messages in this thread | | | Date | Fri, 15 May 2009 11:35:55 +0200 | From | Jens Axboe <> | Subject | Re: do we really want to export more pdflush details in sysctls |
| |
On Wed, May 13 2009, Peter W. Morreale wrote: > On Wed, 2009-05-13 at 15:08 +0200, Jens Axboe wrote: > > On Wed, May 13 2009, Christoph Hellwig wrote: > > > Hi all, > > > > > > commit fafd688e4c0c34da0f3de909881117d374e4c7af titled > > > "mm: add /proc controls for pdflush threads" adds two more sysctl > > > variables exposing details about pdflush threads. At the same time > > > Jens Axboe is working on the per-bdi writeback patchset which will > > > hopefull soon get rid of the pdflush threads in their current form. > > > > > > Is it really a good idea to expose more details now or should we revert > > > this patch before 2.6.30 is out? > > > > Pained me as well when updating the patchset. I see little value in > > these knobs as it is, I'm imagining that the submitter must have had a > > use case where it made some difference? > > > > No, I didn't. The rational was as explained in the commit log, merely > that one size (eg: 2-8 threads) didn't fit all cases, so give the admin > a chance at tuning w/o having to recompile.
OK. In general I think it's a pretty bad idea to add such knobs before there are specific use cases, as we have to maintain them forever. I didn't track where this patch came from, I just spotted it in mainline during the merge window.
> More importantly, I didn't know that Jens was working on significant > changes to writeback. This is sorely needed as from what I see in the > code, writeback is very unfair to 'fast' block devices (when both 'fast' > and 'slow' devices co-exist), and consequently, the apps that reference > them. > > Jens: When do you expect to complete the per-bdi patchset?
Sooner rather than later. I've been working on it the past few days, I needed to make some fundemental changes to support WB_SYNC_ALL and sync(1) properly, unfortunately. I'll be posting an updated patchset early next week.
> In any event, it is not a good idea to expose knobs that will soon be > obviated so please pull the patch.
Good, I have reverted the commit in my for-linus branch and will be asking Linus to pull that soonish.
-- Jens Axboe
| |