On Mon, Dec 11, 2017 at 4:31 PM, Chengguang Xu <cgxu519@xxxxxxxxxx> wrote: > >> 在 2017年12月11日，下午9:20，Chengguang Xu <cgxu519@xxxxxxxxxx> 写道： >> >>> >>> 在 2017年12月11日，下午8:44，Amir Goldstein <amir73il@xxxxxxxxx> 写道： >>> >>> On Mon, Dec 11, 2017 at 2:33 PM, Chengguang Xu <cgxu519@xxxxxxxxxx> wrote: >>>>> >>>>> 在 2017年12月11日，下午6:46，Amir Goldstein <amir73il@xxxxxxxxx> 写道： >>>>> >>>>> On Mon, Dec 11, 2017 at 12:03 PM, Chengguang Xu <cgxu519@xxxxxxxxxx> wrote: >>>>>>> >>>>>>> 在 2017年12月7日，下午4:17，Amir Goldstein <amir73il@xxxxxxxxx> 写道： >>>>> [...] >>>>>> >>>>>> I did more detail tests for three different data modes of ext4 and found >>>>>> the overlayfs syncfs bug is reproducible on data=ordered and data=writeback, >>>>>> but on data=journal mode, data is flushed and correct. >>>>> >>>>> That is expected, because overlayfs does call upper's sync_fs() method and >>>>> for journal=data that will flush all dirty pages as well. >>>>> >>>>>> I only wrote only a few words >>>>>> to a single file and the bug is always reproducible on my test environment. >>>>>> >>>>>> For writeback interferences, AFAIK, from dirty ratio and period. >>>>>> If we drop all dirty caches & sync before the test, I think we can >>>>>> avoid interference from it. >>>>>> >>>>> >>>>> Why? does either drop_caches or sync() reset the flusher thread >>>>> periodic flush dirty pages? >>>> >>>> Sorry，We have to finish test in 30 seconds after we write test file, >>>> otherwise may be affected by background flusher. >>>> >>>> >>>>> >>>>>> So if we don’t have anything else to interference test result, >>>>>> I just want to modify to write a small single file as test target. >>>>>> >>>>>> Am I missing anything? >>>>> >>>>> I think the chance of flusher thread interfering the test and >>>>> flushing the dirty page you wrote before _scratch_shutdown exists, >>>>> but is small enough so we can neglect it and keep the test as simple >>>>> as possible. >>>> >>>> As I know, in normal case flusher thread check dirty inode expiring every 5 seconds(default) and flush dirty >>>> inode when expires 30 seconds(default). If we can finish test in 30s after running test, it would be OK. >>>> What do you think? >>>> >>> >>> 5 seconds is ext4 default journal commit interval. this is when dirty >>> metadata will be flushed. >>> 30 is generic flusher thread interval. >>> What if test started 29 seconds after last flush? >> >> Maybe we have misunderstanding about the flushing intervals. >> I’m saying the intervals which are under /proc/sys/vm to control >> writeback behaviors, it seems not related to any specific filesystem >> like ext4 or others. >> >> dirty_expire_interval >> default:3000(millisecond) >> >> dirty_writeback_interval >> default:500(millisecond) > > Correction: > > dirty_expire_interval > default:3000(centiseconds) > > dirty_writeback_interval > default:500(centiseconds) > > Right. I was confusing the 2 different tunables. I agree your test should be fine with simple small write after sync in an isolated test environment. Amir. -- To unsubscribe from this list: send the line "unsubscribe linux-unionfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
