From: Martin Wilck <[email protected]>
This is a spring-off of the previous discussion under the subject
"libmultipath: fix NULL dereference in get_be64".
Repeating part of the text of my last post there:
pp->uid_attribute may be set from the hwtable, which
means that vendor/product must be known, which in turn means that
uid_attribute can't be set correctly before sysfs_pathinfo() is run.
Thus, in order to be consistent, we need to move the filter_property()
further down in pathinfo(), after the call to sysfs_pathinfo().
However, that has a side effect. As Ben already stated,
pp->uid_attribute currently is *never* set the first time the call
chain pathinfo()->filter_property() is run. After the proposed change,
it would *always* be set in this call chain, possibly leading to more
cases where pathinfo returns PATHINFO_SKIPPED.
This matters e.g. for "multipath -w". When we remove a WWID, we must
be sure get_refwwid() fills in the wwid, which won't happen if
PATHINFO_SKIPPED is returned.
(While looking at this, I discovered that it didn't work in current
upstream either. I guess it has been broken since 0.6.0, or
355291b ("libmultipath: filter for missing property in get_refwwid()"
This lead to the addition of patch 3/3).
In order not to break such use cases, we need to make the
filter_property() test in pathinfo() dependent on DI_BLACKLIST. That
would make a lot of sense, but it'd cause a semantic change.
Comments welcome.
Regards
Martin
Martin Wilck (3):
pathinfo: call filter_property() after sysfs_pathinfo()
libmultipath: pathinfo: call filter_property only with DI_BLACKLIST
multipath -w: allow removing blacklisted paths
libmultipath/configure.c | 4 +-
libmultipath/discovery.c | 19 +++++++--
libmultipath/valid.c | 4 --
tests/Makefile | 2 +-
tests/test-lib.c | 18 ++++----
tests/valid.c | 91 ++++++++++++++++++++++++++++++++++++----
6 files changed, 111 insertions(+), 27 deletions(-)
--
2.29.2
--
dm-devel mailing list
[email protected]
https://www.redhat.com/mailman/listinfo/dm-devel
