From: xie xingguo <xie.xingguo@zte.com.cn>
Date: Sat, 1 Dec 2018 09:42:01 +0000 (+0800)
Subject: osd/OSDMap: fix upmap mis-killing for erasure-coded PGs
X-Git-Tag: v14.1.0~722^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=refs%2Fpull%2F25365%2Fhead;p=ceph.git

osd/OSDMap: fix upmap mis-killing for erasure-coded PGs

The up-set of erasure-coded PGs may include CRUSH_ITEM_NONE,
which as a result causes mis-killing of valid upmap items.

Fixes: https://tracker.ceph.com/issues/37493
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
Signed-off-by: ningtao <ningtao@sangfor.com.cn>
---

diff --git a/src/osd/OSDMap.cc b/src/osd/OSDMap.cc
index a6037b7d95c9..4367ea39c554 100644
--- a/src/osd/OSDMap.cc
+++ b/src/osd/OSDMap.cc
@@ -1790,6 +1790,9 @@ void OSDMap::maybe_remove_pg_upmaps(CephContext *cct,
     nextmap.pg_to_raw_up(pg, &raw, &primary);
     set<int> parents;
     for (auto osd : raw) {
+      // skip non-existent/down osd for erasure-coded PGs
+      if (osd == CRUSH_ITEM_NONE)
+        continue;
       if (type > 0) {
         auto parent = nextmap.crush->get_parent_of_type(osd, type, crush_rule);
         if (parent < 0) {
diff --git a/src/test/osd/TestOSDMap.cc b/src/test/osd/TestOSDMap.cc
index e4b70b365f2f..c25434f92765 100644
--- a/src/test/osd/TestOSDMap.cc
+++ b/src/test/osd/TestOSDMap.cc
@@ -636,6 +636,59 @@ TEST_F(OSDMapTest, CleanPGUpmaps) {
     ASSERT_TRUE(!nextmap.have_pg_upmaps(pgid));
   }
 
+  {
+    // https://tracker.ceph.com/issues/37493
+    pg_t ec_pg(0, my_ec_pool);
+    pg_t ec_pgid = osdmap.raw_pg_to_pg(ec_pg);
+    OSDMap tmpmap; // use a tmpmap here, so we do not dirty origin map..
+    int from = -1;
+    int to = -1;
+    {
+      // insert a valid pg_upmap_item
+      vector<int> ec_up;
+      int ec_up_primary;
+      osdmap.pg_to_raw_up(ec_pgid, &ec_up, &ec_up_primary);
+      ASSERT_TRUE(!ec_up.empty());
+      from = *(ec_up.begin());
+      ASSERT_TRUE(from >= 0);
+      for (int i = 0; i < (int)get_num_osds(); i++) {
+        if (std::find(ec_up.begin(), ec_up.end(), i) == ec_up.end()) {
+          to = i;
+          break;
+        }
+      }
+      ASSERT_TRUE(to >= 0);
+      ASSERT_TRUE(from != to);
+      vector<pair<int32_t,int32_t>> new_pg_upmap_items;
+      new_pg_upmap_items.push_back(make_pair(from, to));
+      OSDMap::Incremental pending_inc(osdmap.get_epoch() + 1);
+      pending_inc.new_pg_upmap_items[ec_pgid] =
+      mempool::osdmap::vector<pair<int32_t,int32_t>>(
+        new_pg_upmap_items.begin(), new_pg_upmap_items.end());
+      tmpmap.deepish_copy_from(osdmap);
+      tmpmap.apply_incremental(pending_inc);
+      ASSERT_TRUE(tmpmap.have_pg_upmaps(ec_pgid));
+    }
+    {
+      // mark one of the target OSDs of the above pg_upmap_item as down
+      OSDMap::Incremental pending_inc(tmpmap.get_epoch() + 1);
+      pending_inc.new_state[to] = CEPH_OSD_UP;
+      tmpmap.apply_incremental(pending_inc);
+      ASSERT_TRUE(!tmpmap.is_up(to));
+      ASSERT_TRUE(tmpmap.have_pg_upmaps(ec_pgid));
+    }
+    {
+      // confirm *maybe_remove_pg_upmaps* won't do anything bad
+      OSDMap::Incremental pending_inc(tmpmap.get_epoch() + 1);
+      OSDMap nextmap;
+      nextmap.deepish_copy_from(tmpmap);
+      nextmap.maybe_remove_pg_upmaps(g_ceph_context, tmpmap,
+        nextmap, &pending_inc);
+      tmpmap.apply_incremental(pending_inc);
+      ASSERT_TRUE(tmpmap.have_pg_upmaps(ec_pgid));
+    }
+  }
+
   {
     // TEST pg_upmap
     {