cluster/afr: Fix gfid unsplit code when renamed filename exceeds NAME_MAX

Summary:
- We noticed some folks name their files all the way up to NAME_MAX (usually 255) and when split-brain is encountered, we fail to heal the file.
- This diff puts an upper bound on the number of bytes we will snprintf into the buffer so that we do not fail the rename.
- This is a port of D3646254 to 3.8

Test Plan: Prove test -- can show it fails without patch as well.

Reviewers: #posix_storage, rwareing

Reviewed By: rwareing

Change-Id: I51c6b28374d4a3f21e29044cb727b4b1da7b69e1
Reviewed-on: https://review.gluster.org/18194
Reviewed-by: Shreyas Siravara <sshreyas@fb.com>
CentOS-regression: Gluster Build System <jenkins@build.gluster.org>
Smoke: Gluster Build System <jenkins@build.gluster.org>

2 files changed, 88 insertions, 4 deletions

diff --git a/tests/basic/afr/gfid-unsplit-shd.t b/tests/basic/afr/gfid-unsplit-shd.t
index 77da5243724..760ea3e0868 100644
--- a/tests/basic/afr/gfid-unsplit-shd.t
+++ b/tests/basic/afr/gfid-unsplit-shd.t
@@ -17,9 +17,6 @@ TEST $CLI volume set $V0 nfs.disable off
 TEST $CLI volume set $V0 cluster.quorum-type none
 TEST $CLI volume set $V0 cluster.heal-timeout 5
 TEST $CLI volume set $V0 cluster.favorite-child-policy majority
-#EST $CLI volume set $V0 cluster.favorite-child-by-majority off
-#EST $CLI volume set $V0 cluster.favorite-child-by-mtime on
-#EST $CLI volume set $V0 cluster.favorite-child-by-size off
 TEST $CLI volume set $V0 cluster.metadata-self-heal off
 TEST $CLI volume set $V0 cluster.data-self-heal off
 TEST $CLI volume set $V0 cluster.entry-self-heal off
@@ -95,4 +92,78 @@ TEST [ "$MD5" == "$HEALED_MD5" ]
 TEST rm -f foo/splitfile
 cd ~
 
+umount $M0
+
+# Part II: Absurdly longfile test
+LONG_FILE_NAME="aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+TEST glusterfs --volfile-id=/$V0 --volfile-server=$H0 $M0 \
+  --attribute-timeout=0 --entry-timeout=0
+
+cd $M0
+mkdir foo
+dd if=/dev/urandom of=foo/${LONG_FILE_NAME} bs=128k count=5 2>/dev/null
+
+MD5=$(md5sum foo/${LONG_FILE_NAME} | cut -d\  -f1)
+
+sleep 1
+cd ~
+
+GFID_PARENT_RAW=$(getfattr -n trusted.gfid -e hex $B0/${V0}1/foo 2>/dev/null | grep trusted.gfid | cut -d= -f2)
+GFID_PARENT_FORMATTED=$(echo "$GFID_PARENT_RAW" | awk '{print substr($1,3,8)"-"substr($1,11,4)"-"substr($1,15,4)"-"substr($1,19,4)"-"substr($1,23,12)}')
+GFID_RAW=$(getfattr -n trusted.gfid -e hex $B0/${V0}1/foo/${LONG_FILE_NAME} 2>/dev/null | grep trusted.gfid | cut -d= -f2)
+GFID_FORMATTED=$(echo "$GFID_RAW" | awk '{print substr($1,3,8)"-"substr($1,11,4)"-"substr($1,15,4)"-"substr($1,19,4)"-"substr($1,23,12)}')
+GFID_LINK_B1="$B0/${V0}1/.glusterfs/$(echo $GFID_RAW | awk '{print substr($0,3,2)"/"substr($0,5,2)"/"substr($1,3,8)"-"substr($1,11,4)"-"substr($1,15,4)"-"substr($1,19,4)"-"substr($1,23,12)}')"
+
+# Create a split-brain by downing a brick, and flipping the
+# gfid on the down brick, then bring the brick back up.
+
+# For good measure kill the first brick so the inode cache is wiped, we don't
+# want any funny business
+TEST kill_brick $V0 $H0 $B0/${V0}1
+TEST $CLI volume start $V0 force
+pkill -f gluster/glustershd
+
+rm -f $GFID_LINK_B1
+TEST setfattr -n "trusted.gfid" -v "0xfd551a5cfddd4c1aa4d096ef09ef5c08" $B0/${V0}1/foo/${LONG_FILE_NAME}
+sleep 1
+TEST touch $B0/${V0}1/foo/${LONG_FILE_NAME}
+
+mkdir -p $B0/${V0}1/.glusterfs/fd/55
+ln $B0/${V0}1/foo/${LONG_FILE_NAME} $B0/${V0}1/.glusterfs/fd/55/fd551a5c-fddd-4c1a-a4d0-96ef09ef5c08
+cd ~
+
+touch $B0/${V0}3/.glusterfs/indices/xattrop/$GFID_FORMATTED
+touch $B0/${V0}3/.glusterfs/indices/xattrop/$GFID_PARENT_FORMATTED
+
+TEST $CLI volume start $V0 force
+EXPECT_WITHIN 20 "1" afr_child_up_status $V0 0
+sleep 5
+
+EXPECT_WITHIN 60 "0" get_pending_heal_count $V0
+
+TEST stat $B0/${V0}1/foo/${LONG_FILE_NAME}
+
+cd $M0
+
+# Tickle the file to trigger the gfid unsplit
+TEST stat foo/${LONG_FILE_NAME}
+sleep 1
+
+# Verify the file is readable
+TEST dd if=foo/${LONG_FILE_NAME} of=/dev/null 2>/dev/null
+
+# Verify entry healing happened on the back-end regardless of the
+# gfid-splitbrain state of the directory.
+TEST stat $B0/${V0}1/foo/${LONG_FILE_NAME}
+
+# Verify the MD5 signature of the file
+HEALED_MD5=$(md5sum foo/${LONG_FILE_NAME} | cut -d\  -f1)
+TEST [ "$MD5" == "$HEALED_MD5" ]
+
+# Verify the file can be removed
+TEST rm -f foo/${LONG_FILE_NAME}
+cd ~
+
+
+
 cleanup
diff --git a/xlators/cluster/afr/src/afr-self-heal-name.c b/xlators/cluster/afr/src/afr-self-heal-name.c
index 9ca56f8bd9d..8e5546a702f 100644
--- a/xlators/cluster/afr/src/afr-self-heal-name.c
+++ b/xlators/cluster/afr/src/afr-self-heal-name.c
@@ -84,6 +84,19 @@ _afr_sh_create_unsplit_loc (struct afr_reply *replies, const int child_idx,
                 ret = ENOMEM;
                 goto err;
         }
+
+        /*
+         * Attempting to do a rename on a file that is too long
+         * will return ENAMETOOLONG and heals will fail. Therefore
+         * we should just limit the name to NAME_MAX
+         * knowing that the filename will be truncated.
+         */
+        if (new_name_len > NAME_MAX) {
+                gf_log ("afr", GF_LOG_WARNING,
+                        "Filename too long, truncating to %d bytes.", NAME_MAX);
+                new_name_len = NAME_MAX;
+        }
+
         snprintf (new_name, new_name_len, ".unsplit_%s_%s", tmp_gfid_str,
                 loc->name);
         unsplit_loc->name = new_name;
@@ -376,7 +389,7 @@ __afr_selfheal_assign_gfid (xlator_t *this, inode_t *parent, uuid_t pargfid,
                                 continue;
                         }
                         gf_log (this->name, GF_LOG_WARNING,
-                                "type[%d] = %d (not %d)", i, 
+                                "type[%d] = %d (not %d)", i,
                                 replies[i].poststat.ia_type, ia_type);
                         if (ia_type != IA_INVAL) {
                                 ret = -EIO;