bug fix: osd: do not cache unused buffer in attrs

attrs only reference the origin bufferlist (decode from MOSDPGPush or
ECSubReadReply message) whose size is much greater than attrs in recovery.
If obc cache it (get_obc maybe cache the attr), this causes the whole origin
bufferlist would not be free until obc is evicted from obc cache. So rebuild
the bufferlist before cache it.

Fixes: #12565
Signed-off-by: Ning Yao <zay11022@gmail.com>
Signed-off-by: Xinze Chi <xmdxcxz@gmail.com>

diff --git a/src/osd/ECBackend.cc b/src/osd/ECBackend.cc
index 47f326efaf17679b3b08f5156b2ee4908cf9d9a4..0899bc86c1c82ebb9dc9eefc8f87472be0c771cb 100644 (file)
--- a/src/osd/ECBackend.cc
+++ b/src/osd/ECBackend.cc
@@ -372,6 +372,15 @@ void ECBackend::handle_recovery_read_complete(
     op.xattrs.swap(*attrs);
 
     if (!op.obc) {
+      // attrs only reference the origin bufferlist (decode from ECSubReadReply message)
+      // whose size is much greater than attrs in recovery. If obc cache it (get_obc maybe
+      // cache the attr), this causes the whole origin bufferlist would not be free until
+      // obc is evicted from obc cache. So rebuild the bufferlist before cache it.
+      for (map<string, bufferlist>::iterator it = op.xattrs.begin();
+           it != op.xattrs.end();
+           ++it) {
+        it->second.rebuild();
+      }
       op.obc = get_parent()->get_obc(hoid, op.xattrs);
       op.recovery_info.size = op.obc->obs.oi.size;
       op.recovery_info.oi = op.obc->obs.oi;

diff --git a/src/osd/ReplicatedBackend.cc b/src/osd/ReplicatedBackend.cc
index 5ddc9fd311efce061cc5109525e367fafe077411..e1cf238b2b42a7db0bacda9aa7b4da265e794f77 100644 (file)
--- a/src/osd/ReplicatedBackend.cc
+++ b/src/osd/ReplicatedBackend.cc
@@ -1811,6 +1811,15 @@ bool ReplicatedBackend::handle_pull_response(
 
   bool first = pi.recovery_progress.first;
   if (first) {
+    // attrs only reference the origin bufferlist (decode from MOSDPGPush message)
+    // whose size is much greater than attrs in recovery. If obc cache it (get_obc maybe
+    // cache the attr), this causes the whole origin bufferlist would not be free until
+    // obc is evicted from obc cache. So rebuild the bufferlist before cache it.
+    for (map<string, bufferlist>::iterator it = pop.attrset.begin();
+         it != pop.attrset.end();
+         ++it) {
+      it->second.rebuild();
+    }
     pi.obc = get_parent()->get_obc(pi.recovery_info.soid, pop.attrset);
     pi.recovery_info.oi = pi.obc->obs.oi;
     pi.recovery_info = recalc_subsets(pi.recovery_info, pi.obc->ssc);