[dm-devel] [PATCH 8/9] libmutipath: continue to use old state on PATH_PENDING

[Benjamin Marzinski]

When pathinfo() sets pp->state to PATH_PENDING, it can cause problems
with path checking.  It should act more like check_path(). When
check_path() sees a new state of PATH_PENDING, it doesn't update the
path state at all, so a path's old state is normally never PATH_PENDING.

As and example of the problems of setting a path to PATH_PENDING, If
check_path() sets a path's state to PATH_UP, then a call to pathinfo()
sets the state to PATH_PENDING, and then another call the check_path()
sets the state to PATH_DOWN, multipathd won't fail the path in the
kernel. Also, if a path's state is PATH_PENDING, and nr_active is
recalculated, that path will count as down, even if the state was
previously PATH_UP. If a path already has a state of PATH_WILD or
PATH_UNCHECKED, changing it to PATH_PENDING won't hurt anything, and it
will help anyone who sees it know what's actually happening. But
otherwise, pathinfo() should leave the previous state alone.

Signed-off-by: Benjamin Marzinski <bmarzins at redhat.com>
---
 libmultipath/discovery.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/libmultipath/discovery.c b/libmultipath/discovery.c
index 729bcb9..d3585f9 100644
--- a/libmultipath/discovery.c
+++ b/libmultipath/discovery.c
@@ -1946,8 +1946,11 @@ int pathinfo(struct path *pp, struct config *conf, int mask)
 
 	if (mask & DI_CHECKER) {
 		if (path_state == PATH_UP) {
-			pp->chkrstate = pp->state = get_state(pp, conf, 0,
-							      path_state);
+			int newstate = get_state(pp, conf, 0, path_state);
+			if (newstate != PATH_PENDING ||
+			    pp->state == PATH_UNCHECKED ||
+			    pp->state == PATH_WILD)
+				pp->chkrstate = pp->state = newstate;
 			if (pp->state == PATH_TIMEOUT)
 				pp->state = PATH_DOWN;
 			if (pp->state == PATH_UP && !pp->size) {
-- 
2.17.2