Bug#35848: UUID() returns UUIDs with the wrong time

offset for time part in UUIDs was 1/1000 of what it
should be. In other words, offset was off.

Also handle the case where we count into the future
when several UUIDs are generated in one "tick", and
then the next call is late enough for us to unwind
some but not all of those borrowed ticks.

Lastly, handle the case where we keep borrowing and
borrowing until the tick-counter overflows by also
changing into a new "numberspace" by creating a new
random suffix.

3 files changed, 86 insertions, 12 deletions

diff --git a/mysql-test/r/func_misc.result b/mysql-test/r/func_misc.result
index c35c2eaccd7..a3df23138e8 100644
--- a/mysql-test/r/func_misc.result
+++ b/mysql-test/r/func_misc.result
@@ -310,6 +310,20 @@ drop table t1;
 SELECT NAME_CONST('var', 'value') COLLATE latin1_general_cs;
 NAME_CONST('var', 'value') COLLATE latin1_general_cs
 value
+select @@session.time_zone into @save_tz;
+set @@session.time_zone='UTC';
+select uuid() into @my_uuid;
+select mid(@my_uuid,15,1);
+mid(@my_uuid,15,1)
+1
+select 24 * 60 * 60 * 1000 * 1000 * 10 into @my_uuid_one_day;
+select concat('0',mid(@my_uuid,16,3),mid(@my_uuid,10,4),left(@my_uuid,8)) into @my_uuidate;
+select floor(conv(@my_uuidate,16,10)/@my_uuid_one_day) into @my_uuid_date;
+select 141427 + datediff(curdate(),'1970-01-01') into @my_uuid_synthetic;
+select @my_uuid_date - @my_uuid_synthetic;
+@my_uuid_date - @my_uuid_synthetic
+0
+set @@session.time_zone=@save_tz;
 End of 5.0 tests
 select connection_id() > 0;
 connection_id() > 0
diff --git a/mysql-test/t/func_misc.test b/mysql-test/t/func_misc.test
index 8e995c2da9c..61bf285c364 100644
--- a/mysql-test/t/func_misc.test
+++ b/mysql-test/t/func_misc.test
@@ -426,6 +426,25 @@ drop table t1;
 #
 SELECT NAME_CONST('var', 'value') COLLATE latin1_general_cs;
 
+#
+# Bug #35848: UUID() returns UUIDs with the wrong time
+#
+select @@session.time_zone into @save_tz;
+
+# make sure all times are UTC so the DayNr won't differ
+set @@session.time_zone='UTC';
+select uuid() into @my_uuid;
+# if version nibble isn't 1, the following calculations will be rubbish
+select mid(@my_uuid,15,1);
+select 24 * 60 * 60 * 1000 * 1000 * 10 into @my_uuid_one_day;
+select concat('0',mid(@my_uuid,16,3),mid(@my_uuid,10,4),left(@my_uuid,8)) into @my_uuidate;
+select floor(conv(@my_uuidate,16,10)/@my_uuid_one_day) into @my_uuid_date;
+select 141427 + datediff(curdate(),'1970-01-01') into @my_uuid_synthetic;
+# these should be identical; date part of UUID should be current date
+select @my_uuid_date - @my_uuid_synthetic;
+
+set @@session.time_zone=@save_tz;
+
 --echo End of 5.0 tests
 
 #
diff --git a/sql/item_strfunc.cc b/sql/item_strfunc.cc
index 56aa44b453e..fbfd21aa006 100644
--- a/sql/item_strfunc.cc
+++ b/sql/item_strfunc.cc
@@ -3370,7 +3370,8 @@ static char clock_seq_and_node_str[]="-0000-000000000000";
   number of 100-nanosecond intervals between
   1582-10-15 00:00:00.00 and 1970-01-01 00:00:00.00.
 */
-#define UUID_TIME_OFFSET ((ulonglong) 141427 * 24 * 60 * 60 * 1000 * 10 )
+#define UUID_TIME_OFFSET ((ulonglong) 141427 * 24 * 60 * 60 * \
+                          1000 * 1000 * 10)
 
 #define UUID_VERSION      0x1000
 #define UUID_VARIANT      0x8000
@@ -3429,24 +3430,64 @@ String *Item_func_uuid::val_str(String *str)
     set_clock_seq_str();
   }
 
-  ulonglong tv=my_getsystime() + UUID_TIME_OFFSET + nanoseq;
-  if (unlikely(tv < uuid_time))
-    set_clock_seq_str();
-  else if (unlikely(tv == uuid_time))
+  ulonglong tv= my_getsystime() + UUID_TIME_OFFSET + nanoseq;
+
+  if (likely(tv > uuid_time))
   {
-    /* special protection from low-res system clocks */
-    nanoseq++;
-    tv++;
+    /*
+      Current time is ahead of last timestamp, as it should be.
+      If we "borrowed time", give it back, just as long as we
+      stay ahead of the previous timestamp.
+    */
+    if (nanoseq)
+    {
+      DBUG_ASSERT((tv > uuid_time) && (nanoseq > 0));
+      /*
+        -1 so we won't make tv= uuid_time for nanoseq >= (tv - uuid_time)
+      */
+      long delta= min(nanoseq, tv - uuid_time -1);
+      tv-= delta;
+      nanoseq-= delta;
+    }
   }
   else
   {
-    if (nanoseq)
+    if (unlikely(tv == uuid_time))
     {
-      tv-=nanoseq;
-      nanoseq=0;
+      /*
+        For low-res system clocks. If several requests for UUIDs
+        end up on the same tick, we add a nano-second to make them
+        different.
+        ( current_timestamp + nanoseq * calls_in_this_period )
+        may end up > next_timestamp; this is OK. Nonetheless, we'll
+        try to unwind nanoseq when we get a chance to.
+        If nanoseq overflows, we'll start over with a new numberspace
+        (so the if() below is needed so we can avoid the ++tv and thus
+        match the follow-up if() if nanoseq overflows!).
+      */
+      if (likely(++nanoseq))
+        ++tv;
+    }
+
+    if (unlikely(tv <= uuid_time))
+    {
+      /*
+        If the admin changes the system clock (or due to Daylight
+        Saving Time), the system clock may be turned *back* so we
+        go through a period once more for which we already gave out
+        UUIDs.  To avoid duplicate UUIDs despite potentially identical
+        times, we make a new random component.
+        We also come here if the nanoseq "borrowing" overflows.
+        In either case, we throw away any nanoseq borrowing since it's
+        irrelevant in the new numberspace.
+      */
+      set_clock_seq_str();
+      tv= my_getsystime() + UUID_TIME_OFFSET;
+      nanoseq= 0;
+      DBUG_PRINT("uuid",("making new numberspace"));
     }
-    DBUG_ASSERT(tv > uuid_time);
   }
+
   uuid_time=tv;
   pthread_mutex_unlock(&LOCK_uuid_generator);