On Sat, 2022-12-03 at 00:43 +0100, mwilck@xxxxxxxx wrote: > From: Martin Wilck <mwilck@xxxxxxxx> > > To check whether we will be able to add a given device can be part > of a multipath map, we have two tests in check_path_valid(): > released_to_systemd() and the O_EXCL test. The former isn't helpful > if "multipath -u" is called for the first time for a given device, > and the latter is only used in the "find_multipaths smart" case, > because > actively opening the device with O_EXCL, even for a very short time, > is prone > to races with other processes. > > It turns out that this may cause issues in some scenarios. We saw > problems in > once case where "find_multipaths greedy" was used with a single > non-multipahted root disk and a very large number of multipath LUNs. > The root disk would first be classified as multipath device. > multipathd > would try to create a map, fail (because the disk was mounted) and > trigger another uevent. But because of the very large number of > multipath > devices, this event was queued up behind thousands of other events, > and > the root device timed out eventually. > > While a simple workaround for the given problem would be proper > blacklisting > or using a different find_multipaths mode, I am proposing a different > solution here. An additional test is added in is_path_valid() which > checks whether the given device is currently in use by 1. sysfs > holders, > 2. mounts (from /proc/self/mountinfo) or 3. swaps (from /proc/swaps). > 2. > and 3. are similar to systemd's device detection after switching > root. > This must not only be done for the device itself, but also for all > its > partitions. For mountinfo and swaps, libmount is utilized. > > With this patch, "multipath -u" will make devices with mounted or > otherwise > used partitions available to systemd early, without waiting for > multipathd > to fail setting up the map and re-triggering an uevent. This should > avoid > the issue described above even without blacklisting. The downside of > it > is a longer runtime of "multipath -u" for almost all devices, in > particular > for real multipath devices. The runtime required for the new checks > was in the > order of 0.1ms-1ms in my tests. Moreover, there is a certain risk > that devices may > wrongly classified as non-multipath because of transient mounts or > holders > created by other processes. > > To make this code compile on older distributions, we need some > additional > checks in create-config.mk. > > Signed-off-by: Martin Wilck <mwilck@xxxxxxxx> Sorry, I forgot: Changes RFC -> v2: - only make the in-use check for "smart" and "greedy" setups (Ben Marzinski) - remove wrong "break" statements (Ben) - add another malloc NULL check (Ben) - Add compatibility checks for older versions of libmount (me) Martin -- dm-devel mailing list dm-devel@xxxxxxxxxx https://listman.redhat.com/mailman/listinfo/dm-devel
