On 2023/3/15 22:14, Martin Wilck wrote: > On Wed, 2023-03-15 at 21:10 +0800, Li Xiao Keng wrote: >>> >> I test move close() after ioctl(). The reason of EBUSY is that >> mdadm >> open(sdf) with O_EXCL. So fd should be closed before ioctl. When I >> remove >> O_EXCL, ioctl() will return success. > > This makes sense. I suppose mdadm must use O_EXCL if it modifies RAID > meta data, otherwise data corruption is just too likely. It is also > impossible to drop the O_EXCL flag with fcntl() without closing the fd. > > So, if mdadm must close the fd before calling ioctl(), the race can > hardly be avoided. The close() will cause a uevent, and nothing > prevents the udev rules from running before the ioctl() returns. > Now I find that close() cause a change udev. Is it necessary to import "mdadm --incremental --export" when change udev cause? Can we ignore it? > Unless we want to explore the flag-in-tmpfs idea, I think mdadm must > expect this to happen, and deal with -EBUSY accordingly. > > But first we should get an answer to Coly's question, which version of > mdadm is in use? > I use 4.1 mdadm and 5.10 kernel. > Martin > > . >
