On Tue, 2007-04-24 at 12:19 +0200, Miklos Szeredi wrote: > > > > > Ahh, now I see; I had totally blocked out these few lines: > > > > > > > > > > pages_written += write_chunk - wbc.nr_to_write; > > > > > if (pages_written >= write_chunk) > > > > > break; /* We've done our duty > > > > > */ > > > > > > > > > > yeah, those look dubious indeed... And reading back Neil's comments, I > > > > > think he agrees. > > > > > > > > > > Shall we just kill those? > > > > > > > > I think we should. > > > > > > > > Athough I'm a little afraid, that Akpm will tell me again, that I'm a > > > > stupid git, and that those lines are in fact vitally important ;) > > > > > > > > > > It depends what they're replaced with. > > > > > > That code is there, iirc, to prevent a process from getting stuck in > > > balance_dirty_pages() forever due to the dirtying activity of other > > > processes. > > > > > > hm, we ask the process to write write_chunk pages each go around the loop. > > > So if it wrote write-chunk/2 pages on the first pass it might end up > > > writing > > > write_chunk*1.5 pages total. I guess that's rare and doesn't matter much > > > if it does happen - the upper bound is write_chunk*2-1, I think. > > > > Right, but I think the problem is that its dirty -> writeback, not dirty > > -> writeback completed. > > > > Ie. they don't guarantee progress, it could be that the total > > nr_reclaimable + nr_writeback will steadily increase due to this break. > > > > How about ensuring that vm_writeout_total increases least > > 2*sync_writeback_pages() during our stay in balance_dirty_pages(). That > > way we have the guarantee that more pages get written out than can be > > dirtied. > > No, because that's a global counter, which many writers could be > looking at. > > We'd need a per-task writeout counter, but when finishing the write we > don't know anymore which task it was performed for.
Yeah, just reached that conclusion myself too - again, I ran into that when trying to figure out how to do the per task balancing right. - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/