On 01/04/2017 04:30 AM, NeilBrown wrote:
>
> This special casing brings no value. It appears to attempt to
> determine if the array is active yet or not, and to skip
> processing if the array has not yet been started.
Hi Neil,
those rules also have another use which is to not trigger further
unnecessary actions if the device is already up and running, hence
avoiding useless resource consumption if it brings no new information.
Currently, this is applied only for running pvscan on top of newly
activated MD device (that's why it's part of 69-dm-lvm-metad.rules
at the moment).
So what we also need is to detect the very first CHANGE event
that makes the device active and to make a difference between this
first CHANGE event and further CHANGE events which are possibly
part of the WATCH rule and any other possible CHANGE events which
do not notify about the device switching from "not ready" to "ready"
state (of course, counting with possible coldplug events).
> However, if the array hasn't been started, then "blkid" will
> not have been able to read a signature, so:
> ENV{ID_FS_TYPE}!="LVM2_member|LVM1_member", GOTO="lvm_end"
> will have caused all this code to be skipped.
>
> Further, this code causes incorrect behaviour in at least one case.
> It assumes that the first "add" event should be ignored, as it will be
> followed by a "change" event which indicates the array coming on line.
> This is consistent with how the kernel sends events, but not always
> consistent with how this script sees event.
> Specifically: if the initrd has "mdadm" support installed, but not
> "lvm2" support, then the initial "add" and "change" events will
> happen while the initrd is in charge and this file is not available.
> Once the root filesystem is mountd, this file will be available
> and "udevadm trigger --action=add" will be run.
> So the first and only event seen by this script for an md device will be
> "add", and it will incorrectly ignore it.
>
Yes, you're right that in this case, it's not behaving correctly when
the initrd doesn't have this rule while the root FS does. To fix this
issue for now, I suggest to separate those rule out of 69-dm-lvm-metad.rules
and make it a part of MD rules so that rule is always available both in
initrd and root fs when MD is used (while LVM doesn't need to be
installed in initrd).
This comes right on time because right at this very moment, I'm working
on a design for a solution which covers this area - I'll surely pass you
the design doc once it's more complete (should be in next few days) so
you we can discuss this problem further. This will cover the *standard*
notification about when the block device is ready which provides a
standard way of letting others (rules or any uevent monitors) know when
the switch from "not ready" to "ready" state happens exactly. This should
save us lots of unnecessary work that is done at the moment - we don't
need to fire scans and further inspection of the device for all the
events all the time. This work also covers identification of spurious
events coming as a result of the WATCH rule and minimization of its
impact on uevent processing performance in userspace.
--
Peter
--
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
