From: Sage Weil <sage@redhat.com>
Date: Tue, 2 May 2017 00:57:45 +0000 (-0500)
Subject: crush/CrushWrapper: remap across intervening bucket types
X-Git-Tag: v12.1.0~10^2~91^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=10ff811643d41940846b94b75b1e635267a14a7a;p=ceph.git

crush/CrushWrapper: remap across intervening bucket types

The previous code could only swap overfull devices with underfull devices
if they were in the same bucket.  With this change, we can map across
buckets.  For example, if host A has more PGs than host B, then we'll
remap some PGs on devices in host A with devices in host B.

Signed-off-by: Sage Weil <sage@redhat.com>
---

diff --git a/src/crush/CrushWrapper.cc b/src/crush/CrushWrapper.cc
index 8fac5f50a7ca..f9eb0ac38e99 100644
--- a/src/crush/CrushWrapper.cc
+++ b/src/crush/CrushWrapper.cc
@@ -2224,6 +2224,29 @@ int CrushWrapper::_choose_type_stack(
   ldout(cct, 10) << __func__ << " cumulative_fanout " << cumulative_fanout
 		 << dendl;
 
+  // identify underful targets for each intermediate level.
+  // this serves two purposes:
+  //   1. we can tell when we are selecting a bucket that does not have any underfull
+  //      devices beneath it.  that means that if the current input includes an overfull
+  //      device, we won't be able to find an underfull device with this parent to
+  //      swap for it.
+  //   2. when we decide we should reject a bucket due to the above, this list gives us
+  //      a list of peers to consider that *do* have underfull devices available..  (we
+  //      are careful to pick one that has the same parent.)
+  vector<set<int>> underfull_buckets; // level -> set of buckets with >0 underfull item(s)
+  underfull_buckets.resize(stack.size() - 1);
+  for (auto osd : underfull) {
+    int item = osd;
+    for (int j = (int)stack.size() - 2; j >= 0; --j) {
+      int type = stack[j].first;
+      item = get_parent_of_type(item, type);
+      ldout(cct, 10) << __func__ << " underfull " << osd << " type " << type
+		     << " is " << item << dendl;
+      underfull_buckets[j].insert(item);
+    }
+  }
+  ldout(cct, 20) << __func__ << " underfull_buckets " << underfull_buckets << dendl;
+
   for (unsigned j = 0; j < stack.size(); ++j) {
     int type = stack[j].first;
     int fanout = stack[j].second;
@@ -2235,17 +2258,22 @@ int CrushWrapper::_choose_type_stack(
     auto tmpi = i;
     for (auto from : w) {
       ldout(cct, 10) << " from " << from << dendl;
-
+      // identify leaves under each choice.  we use this to check whether any of these
+      // leaves are overfull.  (if so, we need to make sure there are underfull candidates
+      // to swap for them.)
+      vector<set<int>> leaves;
+      leaves.resize(fanout);
       for (int pos = 0; pos < fanout; ++pos) {
 	if (type > 0) {
 	  // non-leaf
 	  int item = get_parent_of_type(*tmpi, type);
 	  o.push_back(item);
-	  ldout(cct, 10) << __func__ << "   from " << *tmpi << " got " << item
-			 << " of type " << type << dendl;
 	  int n = cum_fanout;
-	  while (n-- && tmpi != orig.end())
-	    ++tmpi;
+	  while (n-- && tmpi != orig.end()) {
+	    leaves[pos].insert(*tmpi++);
+	  }
+	  ldout(cct, 10) << __func__ << "   from " << *tmpi << " got " << item
+			 << " of type " << type << " over leaves " << leaves[pos] << dendl;
 	} else {
 	  // leaf
 	  bool replaced = false;
@@ -2287,6 +2315,50 @@ int CrushWrapper::_choose_type_stack(
 	  }
 	}
       }
+      if (j + 1 < stack.size()) {
+	// check if any buckets have overfull leaves but no underfull candidates
+	for (int pos = 0; pos < fanout; ++pos) {
+	  if (underfull_buckets[j].count(o[pos]) == 0) {
+	    // are any leaves overfull?
+	    bool any_overfull = false;
+	    for (auto osd : leaves[pos]) {
+	      if (overfull.count(osd)) {
+		any_overfull = true;
+	      }
+	    }
+	    if (any_overfull) {
+	      ldout(cct, 10) << " bucket " << o[pos] << " has no underfull targets and "
+			     << ">0 leaves " << leaves[pos] << " is overfull; alts "
+			     << underfull_buckets[j]
+			     << dendl;
+	      for (auto alt : underfull_buckets[j]) {
+		if (std::find(o.begin(), o.end(), alt) == o.end()) {
+		  // see if alt has the same parent
+		  if (j == 0 ||
+		      get_parent_of_type(o[pos], stack[j-1].first) ==
+		      get_parent_of_type(alt, stack[j-1].first)) {
+		    if (j)
+		      ldout(cct, 10) << "  replacing " << o[pos]
+				     << " (which has no underfull leaves) with " << alt
+				     << " (same parent "
+				     << get_parent_of_type(alt, stack[j-1].first) << " type "
+				     << type << ")" << dendl;
+		    else
+		      ldout(cct, 10) << "  replacing " << o[pos]
+				     << " (which has no underfull leaves) with " << alt
+				     << " (first level)" << dendl;
+		    o[pos] = alt;
+		    break;
+		  } else {
+		    ldout(cct, 30) << "  alt " << alt << " for " << o[pos]
+				   << " has different parent, skipping" << dendl;
+		  }
+		}
+	      }
+	    }
+	  }
+	}
+      }
       if (i == orig.end()) {
 	ldout(cct, 10) << __func__ << " end of orig, break 2" << dendl;
 	break;