Re-enable the statfs cache on the OSTs, unless there isn't much space left,

in order to hopefully remove the perf hit from the 974 changes.

We compensate for potentially fast devices consuming a lot of free space by
slightly over-consuming available space from the cached statfs data as each
BRW is handled.  Not enough to throw out the statfs results by any meaningful
amount (esp. given that we include the slightly-old results from the client
cache when replying to clients).  This overconsumption is thrown out each time
we refresh the statfs data (every second or every RPC if almost full).

DEBUG_REQ() now shows the req->rq_repmsg->status in addition to req->rq_status
since the latter is usually zero and the former has the real reply status.
b=2644

diff --git a/lustre/tests/oos.sh b/lustre/tests/oos.sh
index a780671..dd1c3bb 100755 (executable)
--- a/lustre/tests/oos.sh
+++ b/lustre/tests/oos.sh
@@ -15,6 +15,13 @@ SUCCESS=1
 
 rm -f $OOS
 
+sleep 1        # to ensure we get up-to-date statfs info
+
+#echo -1 > /proc/sys/portals/debug
+#echo 0x40a8 > /proc/sys/portals/subsystem_debug
+#lctl clear
+#lctl debug_daemon start /r/tmp/debug 1024
+
 STRIPECOUNT=`cat /proc/fs/lustre/lov/*/activeobd | head -1`
 ORIGFREE=`cat /proc/fs/lustre/llite/*/kbytesavail | head -1`
 MAXFREE=${MAXFREE:-$((200000 * $STRIPECOUNT))}
@@ -59,6 +66,8 @@ if [ $RECORDSOUT -ne $(($FILESIZE / 1024)) ]; then
         SUCCESS=0
 fi
 
+#lctl debug_daemon stop
+
 rm -f $OOS $LOG
 
 if [ $SUCCESS -eq 1 ]; then