core: glusterfsd keeping fd open in index xlator

Problem: At the time of processing GF_EVENT_PARENT_DOWN
         at brick xlator, it forwards the event to next xlator
         only while xlator ensures no stub is in progress.
         At io-thread xlator it decreases stub_cnt before the process
         a stub and notify EVENT to next xlator

Solution: Introduce a new counter to save stub_cnt and decrease
          the counter after process the stub completely at io-thread
          xlator.
          To avoid brick crash at the time of call xlator_mem_cleanup
          move only brick xlator if detach brick name has found in
          the graph

Note: Thanks to pranith for sharing a simple reproducer to
      reproduce the same

fixes bz#1637934
Change-Id: I1a694a001f7a5417e8771e3adf92c518969b6baa
Signed-off-by: Mohit Agrawal <moagrawal@redhat.com>

4 files changed, 94 insertions, 25 deletions

diff --git a/libglusterfs/src/xlator.c b/libglusterfs/src/xlator.c
index 01cb27896ee..b305ad50e5d 100644
--- a/libglusterfs/src/xlator.c
+++ b/libglusterfs/src/xlator.c
@@ -1051,6 +1051,7 @@ xlator_mem_cleanup(xlator_t *this)
         top = ctx->active->first;
         LOCK(&ctx->volfile_lock);
         /* TODO here we have leak for xlator node in a graph */
+        /* Need to move only top xlator from a graph */
         for (trav_p = &top->children; *trav_p; trav_p = &(*trav_p)->next) {
             victim = (*trav_p)->xlator;
             if (victim->call_cleanup && !strcmp(victim->name, this->name)) {
@@ -1058,19 +1059,6 @@ xlator_mem_cleanup(xlator_t *this)
                 break;
             }
         }
-        /* TODO Sometime brick xlator is not moved from graph so followed below
-           approach to move brick xlator from a graph, will move specific brick
-           xlator from graph only while inode table and mem_acct are cleaned up
-        */
-        trav_p = &top->children;
-        while (*trav_p) {
-            victim = (*trav_p)->xlator;
-            if (victim->call_cleanup && !victim->itable && !victim->mem_acct) {
-                (*trav_p) = (*trav_p)->next;
-            } else {
-                trav_p = &(*trav_p)->next;
-            }
-        }
         UNLOCK(&ctx->volfile_lock);
     }
 }
diff --git a/tests/bugs/glusterd/brick-mux-fd-cleanup.t b/tests/bugs/glusterd/brick-mux-fd-cleanup.t
new file mode 100644
index 00000000000..de11c177b8a
--- /dev/null
+++ b/tests/bugs/glusterd/brick-mux-fd-cleanup.t
@@ -0,0 +1,78 @@
+#!/bin/bash
+
+. $(dirname $0)/../../include.rc
+. $(dirname $0)/../../volume.rc
+
+#This .t tests that the fds from client are closed on brick when gluster volume
+#stop is executed in brick-mux setup.
+
+cleanup;
+TEST glusterd
+TEST pidof glusterd
+
+function keep_fd_open {
+#This function has to be run as background job because opening the fd in
+#foreground and running commands is leading to flush calls on these fds
+#which is making it very difficult to create the race where fds will be left
+#open even after the brick dies.
+    exec 5>$M1/a
+    exec 6>$M1/b
+    while [ -f $M0/a ]; do sleep 1; done
+}
+
+function count_open_files {
+    local brick_pid="$1"
+    local pattern="$2"
+    ls -l /proc/$brick_pid/fd | grep -i "$pattern" | wc -l
+}
+
+TEST $CLI volume set all cluster.brick-multiplex on
+TEST $CLI volume create $V0 replica 2 $H0:$B0/${V0}{0,1}
+TEST $CLI volume create $V1 replica 2 $H0:$B0/${V1}{2,3}
+#Have same configuration on both bricks so that they are multiplexed
+#Delay flush fop for a second
+TEST $CLI volume heal $V0 disable
+TEST $CLI volume heal $V1 disable
+TEST $CLI volume set $V0 delay-gen posix
+TEST $CLI volume set $V0 delay-gen.enable flush
+TEST $CLI volume set $V0 delay-gen.delay-percentage 100
+TEST $CLI volume set $V0 delay-gen.delay-duration 1000000
+TEST $CLI volume set $V1 delay-gen posix
+TEST $CLI volume set $V1 delay-gen.enable flush
+TEST $CLI volume set $V1 delay-gen.delay-percentage 100
+TEST $CLI volume set $V1 delay-gen.delay-duration 1000000
+
+TEST $CLI volume start $V0
+TEST $CLI volume start $V1
+
+TEST $GFS -s $H0 --volfile-id=$V0 --direct-io-mode=enable $M0
+TEST $GFS -s $H0 --volfile-id=$V1 --direct-io-mode=enable $M1
+
+TEST touch $M0/a
+keep_fd_open &
+TEST $CLI volume profile $V1 start
+brick_pid=$(get_brick_pid $V1 $H0 $B0/${V1}2)
+TEST count_open_files $brick_pid "$B0/${V1}2/a"
+TEST count_open_files $brick_pid "$B0/${V1}2/b"
+TEST count_open_files $brick_pid "$B0/${V1}3/a"
+TEST count_open_files $brick_pid "$B0/${V1}3/b"
+
+#If any other flush fops are introduced into the system other than the one at
+#cleanup it interferes with the race, so test for it
+EXPECT "^0$" echo "$($CLI volume profile $V1 info incremental | grep -i flush | wc -l)"
+#Stop the volume
+TEST $CLI volume stop $V1
+
+#Wait for cleanup resources or volume V1
+EXPECT_WITHIN $GRAPH_SWITCH_TIMEOUT "^0$" count_open_files $brick_pid "$B0/${V1}2/a"
+EXPECT_WITHIN $GRAPH_SWITCH_TIMEOUT "^0$" count_open_files $brick_pid "$B0/${V1}2/b"
+EXPECT_WITHIN $GRAPH_SWITCH_TIMEOUT "^0$" count_open_files $brick_pid "$B0/${V1}3/a"
+EXPECT_WITHIN $GRAPH_SWITCH_TIMEOUT "^0$" count_open_files $brick_pid "$B0/${V1}3/b"
+
+TEST rm -f $M0/a #Exit keep_fd_open()
+wait
+
+EXPECT_WITHIN $UMOUNT_TIMEOUT "Y" force_umount $M0
+EXPECT_WITHIN $UMOUNT_TIMEOUT "Y" force_umount $M1
+
+cleanup
diff --git a/xlators/performance/io-threads/src/io-threads.c b/xlators/performance/io-threads/src/io-threads.c
index eac834a6c33..c98d1f6ada4 100644
--- a/xlators/performance/io-threads/src/io-threads.c
+++ b/xlators/performance/io-threads/src/io-threads.c
@@ -123,7 +123,7 @@ __iot_dequeue(iot_conf_t *conf, int *pri)
     if (!stub)
         return NULL;
 
-    GF_ATOMIC_DEC(conf->queue_size);
+    conf->queue_size--;
     conf->queue_sizes[*pri]--;
 
     return stub;
@@ -155,7 +155,8 @@ __iot_enqueue(iot_conf_t *conf, call_stub_t *stub, int pri)
     }
     list_add_tail(&stub->list, &ctx->reqs);
 
-    GF_ATOMIC_INC(conf->queue_size);
+    conf->queue_size++;
+    GF_ATOMIC_INC(conf->stub_cnt);
     conf->queue_sizes[pri]++;
 }
 
@@ -183,7 +184,7 @@ iot_worker(void *data)
                 conf->ac_iot_count[pri]--;
                 pri = -1;
             }
-            while (GF_ATOMIC_GET(conf->queue_size) == 0) {
+            while (conf->queue_size == 0) {
                 if (conf->down) {
                     bye = _gf_true; /*Avoid sleep*/
                     break;
@@ -230,6 +231,7 @@ iot_worker(void *data)
             } else {
                 call_resume(stub);
             }
+            GF_ATOMIC_DEC(conf->stub_cnt);
         }
         stub = NULL;
 
@@ -837,9 +839,9 @@ __iot_workers_scale(iot_conf_t *conf)
                                thread_name);
         if (ret == 0) {
             conf->curr_count++;
-            gf_msg_debug(
-                conf->this->name, 0, "scaled threads to %d (queue_size=%ld/%d)",
-                conf->curr_count, GF_ATOMIC_GET(conf->queue_size), scale);
+            gf_msg_debug(conf->this->name, 0,
+                         "scaled threads to %d (queue_size=%d/%d)",
+                         conf->curr_count, conf->queue_size, scale);
         } else {
             break;
         }
@@ -1231,7 +1233,7 @@ init(xlator_t *this)
     GF_OPTION_INIT("pass-through", this->pass_through, bool, out);
 
     conf->this = this;
-    GF_ATOMIC_INIT(conf->queue_size, 0);
+    GF_ATOMIC_INIT(conf->stub_cnt, 0);
 
     for (i = 0; i < GF_FOP_PRI_MAX; i++) {
         INIT_LIST_HEAD(&conf->clients[i]);
@@ -1282,7 +1284,7 @@ notify(xlator_t *this, int32_t event, void *data, ...)
 {
     iot_conf_t *conf = this->private;
     xlator_t *victim = data;
-    uint64_t queue_size = 0;
+    uint64_t stub_cnt = 0;
     struct timespec sleep_till = {
         0,
     };
@@ -1292,14 +1294,14 @@ notify(xlator_t *this, int32_t event, void *data, ...)
             clock_gettime(CLOCK_REALTIME, &sleep_till);
             sleep_till.tv_sec += 1;
             /* Wait for draining stub from queue before notify PARENT_DOWN */
-            queue_size = GF_ATOMIC_GET(conf->queue_size);
+            stub_cnt = GF_ATOMIC_GET(conf->stub_cnt);
 
             pthread_mutex_lock(&conf->mutex);
             {
-                while (queue_size) {
+                while (stub_cnt) {
                     (void)pthread_cond_timedwait(&conf->cond, &conf->mutex,
                                                  &sleep_till);
-                    queue_size = GF_ATOMIC_GET(conf->queue_size);
+                    stub_cnt = GF_ATOMIC_GET(conf->stub_cnt);
                 }
             }
             pthread_mutex_unlock(&conf->mutex);
diff --git a/xlators/performance/io-threads/src/io-threads.h b/xlators/performance/io-threads/src/io-threads.h
index f80fd41c45d..2980d18e74b 100644
--- a/xlators/performance/io-threads/src/io-threads.h
+++ b/xlators/performance/io-threads/src/io-threads.h
@@ -63,7 +63,8 @@ struct iot_conf {
     int32_t ac_iot_limit[GF_FOP_PRI_MAX];
     int32_t ac_iot_count[GF_FOP_PRI_MAX];
     int queue_sizes[GF_FOP_PRI_MAX];
-    gf_atomic_t queue_size;
+    int32_t queue_size;
+    gf_atomic_t stub_cnt;
     pthread_attr_t w_attr;
     gf_boolean_t least_priority; /*Enable/Disable least-priority */