objecter: avoid race when reset down osd's session

consider the sequence step of follow:
thread A: objecter handle new osdmap which says osd.X down
thread B: objecter reset session of osd.X,then reopen session
thread A: close seesion of osd.X because he is down

in such circumstance, we will hit failure on assert(is_up(osd))
the backtrace is:
1. ceph::__ceph_assert_fail(char const*, char const*, int, char const*) ()
2. OSDMap::get_inst(int) const [clone .part.224] ()
3. Objecter::_reopen_session(Objecter::OSDSession*) ()
4. Objecter::ms_handle_reset(Connection*) ()

Fixes: https://tracker.ceph.com/issues/24601
Signed-off-by: Zengran Zhang <zhangzengran@sangfor.com.cn>

diff --git a/src/osdc/Objecter.cc b/src/osdc/Objecter.cc
index 9f5610770cf7ca036a7ce75b34c127045a3c3780..fb0eea8759e27fe45ad2b51c0dc9f02ea13b307b 100644 (file)
--- a/src/osdc/Objecter.cc
+++ b/src/osdc/Objecter.cc
@@ -4384,7 +4384,10 @@ bool Objecter::ms_handle_reset(Connection *con)
     if (session) {
       ldout(cct, 1) << "ms_handle_reset " << con << " session " << session
                    << " osd." << session->osd << dendl;
-      if (!initialized) {
+      // the session maybe had been closed if new osdmap just handled
+      // says the osd down
+      if (!(initialized && osdmap->is_up(session->osd))) {
+       ldout(cct, 1) << "ms_handle_reset aborted,initialized=" << initialized << dendl;
        wl.unlock();
        return false;
       }