From: xie xingguo <xie.xingguo@zte.com.cn>
Date: Thu, 18 Feb 2016 07:33:19 +0000 (+0800)
Subject: common/TrackedOp: fix inaccurate counting for total slow requests
X-Git-Tag: v10.1.0~246^2~1
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=b2bfd6cc0ba697414a0abca1bb8a1074abb08993;p=ceph.git

common/TrackedOp: fix inaccurate counting for total slow requests

In the original design there are two counters in charge of collecting
potentially problematic requests, namely 'slow' and 'warned'.
Counter 'slow' is responsible for capturing all the requests which have
already hit the "complain" limit and shall be marked as "slow" while
counter 'warned' is responsible for countering those requests which
have already hit the "warning interval" and thus shall be logged and
then outputted.

The problem here is if 'warned' counter hits the log_threshold,
we will quit the entire for loop but there may be residual shard_queues
which may still containing slow requests. As a result, the 'slow' counter
does not reflect the real number of total slow requests in all the
shard_queues under this case. And no slow requests will be tracked
especially when 'log_threshold' is set to zero(Do we do this intentional?
Or else we shall never allow 'log_threshold' to be zero).

The solution for the above problem is to keep counting 'slow' requests
until we have finished traversing all the shard_queues, no matter
whether we have gathered enough requests for logging or not, and if so,
we simply stop counter 'warned' and skip over logging process.

Fixes: #14804
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
---

diff --git a/src/common/TrackedOp.cc b/src/common/TrackedOp.cc
index f76587e7b167..39a13b4c2832 100644
--- a/src/common/TrackedOp.cc
+++ b/src/common/TrackedOp.cc
@@ -203,8 +203,7 @@ bool OpTracker::check_ops_in_flight(std::vector<string> &warning_vector)
 
   int slow = 0;     // total slow
   int warned = 0;   // total logged
-  for (uint32_t iter = 0;
-       iter < num_optracker_shards && warned < log_threshold; iter++) {
+  for (uint32_t iter = 0; iter < num_optracker_shards; iter++) {
     ShardedTrackingData* sdata = sharded_in_flight_list[iter];
     assert(NULL != sdata);
     Mutex::Locker locker(sdata->ops_in_flight_lock_sharded);
@@ -215,12 +214,10 @@ bool OpTracker::check_ops_in_flight(std::vector<string> &warning_vector)
       slow++;
 
       // exponential backoff of warning intervals
-      if (((*i)->get_initiated() +
-	 (complaint_time * (*i)->warn_interval_multiplier)) < now) {
-      // will warn
+      if (warned < log_threshold &&
+         ((*i)->get_initiated() + (complaint_time * (*i)->warn_interval_multiplier)) < now) {
+        // will warn, increase counter
         warned++;
-        if (warned > log_threshold)
-          break;
 
         utime_t age = now - (*i)->get_initiated();
         stringstream ss;