cluster/afr: Update read-child if it becomes stale

Change-Id: I00c714a89575023f6dbdd3430dcbf191e5d08019
BUG: 3650
Reviewed-on: http://review.gluster.com/740
Tested-by: Gluster Build System <jenkins@build.gluster.com>
Reviewed-by: Vijay Bellur <vijay@gluster.com>

3 files changed, 30 insertions, 36 deletions

diff --git a/xlators/cluster/afr/src/afr-common.c b/xlators/cluster/afr/src/afr-common.c
index 108b021b32f..9512f6a5477 100644
--- a/xlators/cluster/afr/src/afr-common.c
+++ b/xlators/cluster/afr/src/afr-common.c
@@ -325,7 +325,6 @@ afr_inode_ctx_set_read_child (afr_inode_ctx_t *ctx, int32_t read_child)
         uint64_t        remaining_mask = 0;
         uint64_t        mask         = 0;
 
-        GF_ASSERT (read_child >= 0);
         remaining_mask = (~AFR_ICTX_READ_CHILD_MASK & ctx->masks);
         mask = (AFR_ICTX_READ_CHILD_MASK & read_child);
         ctx->masks = remaining_mask | mask;
@@ -347,19 +346,23 @@ afr_inode_ctx_set_read_ctx (afr_inode_ctx_t *ctx, int32_t read_child,
 }
 
 void
-afr_inode_ctx_rm_stale_children (afr_inode_ctx_t *ctx, int32_t read_child,
-                                 int32_t *stale_children, int32_t child_count)
+afr_inode_ctx_rm_stale_children (afr_inode_ctx_t *ctx, int32_t *stale_children,
+                                 int32_t child_count)
 {
         int             i            = 0;
+        int32_t         read_child   = -1;
 
         GF_ASSERT (stale_children);
-        afr_inode_ctx_set_read_child (ctx, read_child);
         for (i = 0; i < child_count; i++) {
-                if ((ctx->fresh_children[i] == -1) || (stale_children[i] == -1))
+                if (stale_children[i] == -1)
                         break;
                 afr_children_rm_child (ctx->fresh_children,
                                        stale_children[i], child_count);
         }
+        read_child = (int32_t)(ctx->masks & AFR_ICTX_READ_CHILD_MASK);
+        if (!afr_is_child_present (ctx->fresh_children, child_count,
+                                   read_child))
+                afr_inode_ctx_set_read_child (ctx, ctx->fresh_children[0]);
 }
 
 void
@@ -421,9 +424,8 @@ afr_inode_set_ctx (xlator_t *this, inode_t *inode, afr_inode_params_t *params)
                                                     priv->child_count);
                         break;
                 case AFR_INODE_RM_STALE_CHILDREN:
-                        read_child = params->u.read_ctx.read_child;
                         stale_children = params->u.read_ctx.children;
-                        afr_inode_ctx_rm_stale_children (ctx, read_child,
+                        afr_inode_ctx_rm_stale_children (ctx,
                                                          stale_children,
                                                          priv->child_count);
                         break;
@@ -488,16 +490,14 @@ afr_inode_set_read_ctx (xlator_t *this, inode_t *inode, int32_t read_child,
 }
 
 void
-afr_inode_rm_stale_children (xlator_t *this, inode_t *inode, int32_t read_child,
+afr_inode_rm_stale_children (xlator_t *this, inode_t *inode,
                              int32_t *stale_children)
 {
         afr_inode_params_t params = {0};
 
-        GF_ASSERT (read_child >= 0);
         GF_ASSERT (stale_children);
 
         params.op = AFR_INODE_RM_STALE_CHILDREN;
-        params.u.read_ctx.read_child     = read_child;
         params.u.read_ctx.children = stale_children;
         afr_inode_set_ctx (this, inode, &params);
 }
diff --git a/xlators/cluster/afr/src/afr-transaction.c b/xlators/cluster/afr/src/afr-transaction.c
index d3960dcff40..89fbd9ec754 100644
--- a/xlators/cluster/afr/src/afr-transaction.c
+++ b/xlators/cluster/afr/src/afr-transaction.c
@@ -382,44 +382,38 @@ afr_transaction_rm_stale_children (call_frame_t *frame, xlator_t *this,
         local = frame->local;
         pending = local->pending;
 
-        stale_children = afr_children_create (priv->child_count);
-        if (!stale_children)
+        if (local->op_ret < 0)
                 goto out;
-
         fresh_children = local->fresh_children;
         read_child = afr_inode_get_read_ctx (this, inode, fresh_children);
-
-        GF_ASSERT (read_child >= 0);
-
-        if (pending[read_child][idx] == 0)
-                read_child = -1;
+        if (read_child < 0) {
+                gf_log (this->name, GF_LOG_DEBUG, "Possible split-brain "
+                        "for %s", uuid_utoa (inode->gfid));
+                goto out;
+        }
 
         for (i = 0; i < priv->child_count; i++) {
                 if (!afr_is_child_present (fresh_children,
                                            priv->child_count, i))
                         continue;
-                if (pending[i][idx] == 0) {
-                        /* child is down or op failed on it */
-                        rm_stale_children = _gf_true;
-                        afr_children_rm_child (fresh_children, i,
-                                               priv->child_count);
-                        stale_children[count++] = i;
-                }
-        }
+                if (pending[i][idx])
+                        continue;
+                /* child is down or op failed on it */
+                if (!stale_children)
+                        stale_children = afr_children_create (priv->child_count);
+                if (!stale_children)
+                        goto out;
 
-        if (!rm_stale_children) {
-                GF_ASSERT (read_child >= 0);
-                goto out;
+                rm_stale_children = _gf_true;
+                stale_children[count++] = i;
+                gf_log (this->name, GF_LOG_DEBUG, "Removing stale child "
+                        "%d for %s", i, uuid_utoa (inode->gfid));
         }
 
-        if (fresh_children[0] == -1) {
-                //All children failed. leave as-is
+        if (!rm_stale_children)
                 goto out;
-        }
 
-        if (read_child == -1)
-                read_child = fresh_children[0];
-        afr_inode_rm_stale_children (this, inode, read_child, stale_children);
+        afr_inode_rm_stale_children (this, inode, stale_children);
 out:
         if (stale_children)
                 GF_FREE (stale_children);
diff --git a/xlators/cluster/afr/src/afr.h b/xlators/cluster/afr/src/afr.h
index e7c948dba35..0566495fe96 100644
--- a/xlators/cluster/afr/src/afr.h
+++ b/xlators/cluster/afr/src/afr.h
@@ -970,7 +970,7 @@ int32_t
 afr_resultant_errno_get (int32_t *children,
                          int *child_errno, unsigned int child_count);
 void
-afr_inode_rm_stale_children (xlator_t *this, inode_t *inode, int32_t read_child,
+afr_inode_rm_stale_children (xlator_t *this, inode_t *inode,
                              int32_t *stale_children);
 void
 afr_launch_self_heal (call_frame_t *frame, xlator_t *this, inode_t *inode,