On Wed, Jul 20 2016, Vaughan wrote: > Hi Neil, > > I'm using v3.10 md code for develop. Recently I encounter a problem where an > read IO usually returned from physical disk after md has been stopped. > I reviewed the code and find when md stops, it unregister raid5d > unconditionally and call shrink_stripes() to free only the *inactive* > stripes. > I know before stop, it uses O_EXCL open the md, but that won't stop others > open it and send IO to it. > So I think it's possible that some active stripes will be still running. do_md_stop() calls sync_blockdev() which was supposed to wait for all outstanding IO. It probably doesn't wait for reads though, only writes. > > And I also found > commit 5aa61f427e4979be733e4847b9199ff9cc48a47e > Author: NeilBrown <neilb@xxxxxxx> > Date: Mon Dec 15 12:56:57 2014 +1100 > md: split detach operation out from ->stop. > > add calling a quiesce before unregister raid5d in __md_stop, which not > exists there before. > Does this fix the hole when md stop? Why don't you try it and see? NeilBrown
Attachment:
signature.asc
Description: PGP signature
