osd: ignore info racing with PGRemove

We may get multiple infos during peering, because we explicitly
query priors AND osds with existing pg data send them (thus we
may get 2).

The second info can race with a PGRemove.  When we get it, we have
to realize it's not _new_ stray content.. it's old content we're
already purging.

diff --git a/src/TODO b/src/TODO
index 866b1feb359224de2dae90c2b779b03d88f6bfd0..bc9418a21fabf10e337b23dbef20a5d442b7a129 100644 (file)
--- a/src/TODO
+++ b/src/TODO
@@ -138,6 +138,7 @@ btrfs
 - ioctl to pull out data csum?
 
 osd
+- remove needs to be in a work queue!
 - what to do with lost objects.. continue peering?
 - segregate backlog from log ondisk?
 - preserve pg logs on disk for longer period

diff --git a/src/osd/OSD.cc b/src/osd/OSD.cc
index fce332b6517ce8b5e86e71a619c2ba118b46e7cc..d4e65cd0ae50596fa54f28771c91628e87ecfd5d 100644 (file)
--- a/src/osd/OSD.cc
+++ b/src/osd/OSD.cc
@@ -2934,9 +2934,13 @@ void OSD::handle_pg_notify(MOSDPGNotify *m)
     // stray?
     bool acting = pg->is_acting(from);
     if (!acting) {
-      dout(10) << *pg << " osd" << from << " has stray content: " << *it << dendl;
-      pg->stray_set.insert(from);
-      pg->state_clear(PG_STATE_CLEAN);
+      if (pg->stray_purged.count(from)) {
+       dout(10) << *pg << " osd" << from << " sent racing info " << *it << "; already purging/purged" << dendl;
+      } else {
+       dout(10) << *pg << " osd" << from << " has stray content: " << *it << dendl;
+       pg->stray_set.insert(from);
+       pg->state_clear(PG_STATE_CLEAN);
+      }
     }
 
     if (had) {

diff --git a/src/osd/PG.cc b/src/osd/PG.cc
index 58bcd1357bd04400dd43f01a7cc310affacba6c3..e813da472a90f5d2f084019e6df3b76239cda616 100644 (file)
--- a/src/osd/PG.cc
+++ b/src/osd/PG.cc
@@ -1005,6 +1005,7 @@ void PG::clear_primary_state()
   need_up_thru = false;
   peer_last_complete_ondisk.clear();
   min_last_complete_ondisk = eversion_t();
+  stray_purged.clear();
 
   finish_sync_event = 0;  // so that _finish_recvoery doesn't go off in another thread
 
@@ -1710,6 +1711,7 @@ void PG::purge_strays()
     if (osd->osdmap->is_up(*p)) {
       dout(10) << "sending PGRemove to osd" << *p << dendl;
       osd->queue_for_removal(*p, info.pgid);
+      stray_purged.insert(*p);
     } else {
       dout(10) << "not sending PGRemove to down osd" << *p << dendl;
     }

diff --git a/src/osd/PG.h b/src/osd/PG.h
index 2566b40c70ba2232f697f6b468d3ef5162905125..8c01e2dd6ca5b233f7e3dfc5c2dca5078f1b657f 100644 (file)
--- a/src/osd/PG.h
+++ b/src/osd/PG.h
@@ -699,6 +699,7 @@ public:
   map<int, Missing>    peer_missing;
   set<int>             peer_log_requested;  // logs i've requested (and start stamps)
   set<int>             peer_summary_requested;
+  set<int>             stray_purged;  // i deleted these strays; ignore racing PGInfo from them
   friend class OSD;