Hi Wu, You're right, the BDI threads should be woken up reliably by the balance_dirty_pages() and balance_dirty_pages() needs to be called from all code that is responsible for dirtying the pages. Sorry, I was not too aware of the balance_dirty_pages() functionality and the way it was being called in entirety or I would have spotted this. Thanks for adding the dirty_background_time into your over_bground_thresh() formula. Now that you seem to have included the time into the threshold, I can relate to your patch better as a solution for the problems I earlier mentioned. Thanks again, Kautuk. On Fri, Aug 19, 2011 at 7:54 PM, Wu Fengguang <fengguang.wu@xxxxxxxxx> wrote: > Hi Kautuk, > > On Fri, Aug 19, 2011 at 03:00:30PM +0800, Kautuk Consul wrote: >> Hi Wu, >> >> Yes. I think I do understand your approach. >> >> Your aim is to always retain the per BDI timeout value. >> >> You want to check for threshholds by mathematically adjusting the >> background time too >> into your over_bground_thresh() formula so that your understanding >> holds true always and also >> affects the page dirtying scenario I mentioned. >> This definitely helps and refines this scenario in terms of flushing >> out of the dirty pages. > > Thanks. > >> Doubts: >> i) Your entire implementation seems to be dependent on someone >> calling balance_dirty_pages() >> directly or indirectly. This function will call the >> bdi_start_background_writeback() which wakes >> up the flusher thread. >> What about those page dirtying code paths which might not call >> balance_dirty_pages ? >> Those paths then depend on the BDI thread periodically writing it >> to disk and then we are again >> dependent on the writeback interval. >> Can we assume that the kernel will reliably call >> balance_dirty_pages() whenever the pages >> are dirtied ? If that was true, then we would not need bdi >> periodic writeback threads ever. > > Yes. The kernel need a way to limit the total number of dirty pages at > any given time and to keep them under dirty_ratio/dirty_bytes. > > balance_dirty_pages() is such a central place to throttle the dirty > pages. Whatever code path generating dirty pages are required to call > into balance_dirty_pages_ratelimited_nr() which will in turn call > balance_dirty_pages(). > > So, the values specified by dirty_ratio/dirty_bytes will be executed > effectively by balance_dirty_pages(). In contrast, the values > specified by dirty_expire_centisecs is merely a parameter used by > wb_writeback() to select the eligible inodes to do writeout. The 30s > dirty expire time is never a guarantee that all inodes/pages dirtied > before 30s will be timely written to disk. It's better interpreted in > the opposite way: when under the dirty_background_ratio threshold and > hence background writeout does not kick in, dirty inodes younger than > 30s won't be written to disk by the flusher. > >> ii) Even after your rigorous checking, the bdi_writeback_thread() >> will still do a schedule_timeout() >> with the global value. Will your current solution then handle >> Artem's disk removal scenario ? >> Else, you start using your value in the schedule_timeout() call >> in the bdi_writeback_thread() >> function, which brings us back to the interval phenomenon I was >> talking about. > > wb_writeback() will keep running as long as over_bground_thresh(). > > The flusher will keep writing as long as there are more works, since > there is a > > if (!list_empty(&bdi->work_list)) > continue; > > before the schedule_timeout() call. > > And the flusher thread will always be woke up timely from > balance_dirty_pages(). > > So schedule_timeout() won't block in the way at all. > >> Does this patch really help the user control exact time when the write >> BIO is transferred from the >> MM to the Block layer assuming balance_dirty_pages() is not called ? > > It would be a serious bug if balance_dirty_pages() is somehow not > called. But note that balance_dirty_pages() is designed to be called > on every N pages to reduce overheads. > > Thanks, > Fengguang > -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
