mdadm: enable sync file for udev rules

Mounting an md device may fail during boot from mdadm's claim
on the device not being released before systemd attempts to mount.

In this case it was found that essentially there is a race condition
occurring in which the mount cannot happen without some kind of delay
being added BEFORE the mount itself triggers, or manual intervention
after a timeout.

The findings:
the inode was for a tmp block node made by mdadm for md0.

crash> detailedsearch ff1b0c398ff28380
ff1b0c398f079720: ff1b0c398ff28380 slab:filp state:alloc
 obj:ff1b0c398f079700 size:256
ff1b0c398ff284f8: ff1b0c398ff28380 slab:shmem_inode_cache
 state:alloc obj:ff1b0c398ff28308 size:768

crash> struct file.f_inode,f_path ff1b0c398f079700
f_inode = 0xff1b0c398ff28380,
f_path = {
mnt = 0xff1b0c594aecc7a0,
dentry = 0xff1b0c3a8c614f00
},
crash> struct dentry.d_name 0xff1b0c3a8c614f00
d_name = {
{
{ hash = 3714992780, len = 16 },
hash_len = 72434469516
},
name = 0xff1b0c3a8c614f38 ".tmp.md.1454:9:0"
},

For the race condition, mdadm and udev have some infrastructure for making
the device be ignored while under construction. e.g.

$ cat lib/udev/rules.d/01-md-raid-creating.rules

do not edit this file, it will be overwritten on update
While mdadm is creating an array, it creates a file
/run/mdadm/creating-mdXXX. If that file exists, then
the array is not "ready" and we should make sure the
content is ignored.
KERNEL=="md*", TEST=="/run/mdadm/creating-$kernel", ENV{SYSTEMD_READY}="0"

However, this feature currently is only used by the mdadm create command.
See calls to udev_block/udev_unblock in the mdadm code as to where and when
this behavior is used. Any md array being started by incremental or
normal assemble commands does not use this udev integration. So assembly
of an existing array does not look to have any explicit protection from
systemd/udev seeing an array as in a usable state before an mdadm instance
with O_EXCL closes its file handle.
This is for the sake of showing the use case for such an option and why
it would be helpful to delay the mount itself.

While mdadm is still constructing the array mdadm --incremental
that is called from within /usr/lib/udev/rules.d/64-md-raid-assembly.rules,
there is an attempt to mount the md device, but there is not a creation
of "/run/mdadm/creating-xxx" file when in incremental mode that
the rule is looking for.  Therefore the device is not marked
as SYSTEMD_READY=0  in
"/usr/lib/udev/rules.d/01-md-raid-creating.rules" and missing
synchronization using the "/run/mdadm/creating-xxx" file.

As to this change affecting containers or IMSM...
(container's array state is inactive all the time)

Even if the "array_state" reports "inactive" when previous components
are added, the mdadm call for the very last array component that makes
it usable/ready, still needs to be synced properly - mdadm needs to drop
the claim first calling "close", then delete the "/run/mdadm/creating-xxx".
Then lets the udev know it is clear to act now (the "udev_unblock" in
mdadm code that generates a synthetic udev event so the rules are
reevalutated). It's this processing of the very last array component
that is the issue here (which is not IO error, but it is that trying to
open the dev returns -EBUSY because of the exclusive claim that mdadm
still holds while the mdadm device is being processed already by udev in
parallel, and that is what the
/run/mdadm/creating-xxx should prevent exactly).

The patch to Incremental.c is to enable creating the
"/run/mdadm/creating-xxx" file during incremental mode.

For the change to Create.c, the unlink is called right before dropping
the exculusive claim for the device. This should be the other way round
to avoid the race 100%. That is, if there's a "close" call and
"udev_unblock" call, the "close" should go first, then followed
"udev_unblock".

Signed-off-by: Nigel Croxon <ncroxon@redhat.com>