DAOS-10622 pool: Fix ds_pool_get_version for NULL sp_map

Makito and Samir observed the following assertion failure after
restarting engines.

  #0  raise () from /lib64/libc.so.6
  #1  abort () from /lib64/libc.so.6
  #2  __assert_fail_base () from /lib64/libc.so.6
  #3  __assert_fail () from /lib64/libc.so.6
  #4  pool_map_get_version (map=0x0) at src/common/pool_map.c:2852
  #5  ds_pool_get_version (pool=0x7f0ca063c690, pool=0x7f0ca063c690) at
      src/include/daos_srv/pool.h:296
  #6  pc=rpc@entry=0x7f0ca0998d30, p_rpt=p_rpt@entry=0x7f0ca83a77b0) at
      src/rebuild/srv.c:2101
  #7  rebuild_tgt_scan_handler (rpc=0x7f0ca0998d30) at
      src/rebuild/scan.c:954
  #8  crt_handle_rpc (arg=0x7f0ca0998d30) at src/cart/crt_rpc.c:1654
  #9  ABTD_ythread_func_wrapper (p_arg=0x7f0ca83a78a0) at
      arch/abtd_ythread.c:21
  #10 make_fcontext () from /usr/lib64/libabt.so.1
  #11 ?? ()

The ds_pool_get_version call passed a NULL map argument to
pool_map_get_version. The ds_pool.sp_map field may be NULL after the
pool is started but before the pool receives the initial pool map from
the pool service. This patch fixes ds_pool_get_version to return 0,
which is less than all valid pool map versions, when sp_map is NULL,
resulting in rebuild retries like this:

  Rebuild [queued] (pool=3bf68c9c ver=2) tgts=2
  Rebuild [started] (pool 3bf68c9c ver=2)
  Rebuild [failed] (pool 3bf68c9c ver=2 status=DER_BUSY(-1012): 'Device
    or resource busy')
  Rebuild [queued] (pool=3bf68c9c ver=2) tgts=2
  Rebuild [started] (pool 3bf68c9c ver=2)
  Rebuild [failed] (pool 3bf68c9c ver=2 status=DER_BUSY(-1012): 'Device
    or resource busy')
  Rebuild [queued] (pool=3bf68c9c ver=2) tgts=2
  Rebuild [started] (pool 3bf68c9c ver=2)
  Rebuild [scanning] (pool 3bf68c9c ver=2, toberb_obj=0, rb_obj=0, [...]
  Rebuild [scanning] (pool 3bf68c9c ver=2, toberb_obj=0, rb_obj=0, [...]
  Rebuild [completed] (pool 3bf68c9c ver=2, toberb_obj=0, rb_obj=0,[...]
  Target[2] (rank 2 idx 0 status 16 ver 1) is excluded.

Also, this patch removes some rebuild code that handles NULL
ds_pool.sp_group fields. Those can not happen as we always initialize
sp_group (as well as sp_iv_ns) before putting a ds_pool object into the
LRU.

Signed-off-by: Li Wei <[email protected]>

src/include/daos_srv/pool.h

@@ -297,10 +297,11 @@ ds_pool_rf_verify(struct ds_pool *pool, uint32_t last_ver, uint32_t rf)
 static inline uint32_t
 ds_pool_get_version(struct ds_pool *pool)
 {
-	uint32_t	ver;
+	uint32_t	ver = 0;
 
 	ABT_rwlock_rdlock(pool->sp_lock);
-	ver = pool_map_get_version(pool->sp_map);
+	if (pool->sp_map != NULL)
+		ver = pool_map_get_version(pool->sp_map);
 	ABT_rwlock_unlock(pool->sp_lock);
 
 	return ver;

src/rebuild/srv.c

@@ -2111,19 +2111,9 @@ rebuild_tgt_prepare(crt_rpc_t *rpc, struct rebuild_tgt_pool_tracker **p_rpt)
 		D_GOTO(out, rc = -DER_BUSY);
 	}
 
-	if (pool->sp_group == NULL) {
-		char id[DAOS_UUID_STR_SIZE];
-
-		uuid_unparse_lower(pool->sp_uuid, id);
-		pool->sp_group = crt_group_lookup(id);
-		if (pool->sp_group == NULL) {
-			D_ERROR(DF_UUID": pool group not found\n",
-				DP_UUID(pool->sp_uuid));
-			D_GOTO(out, rc = -DER_INVAL);
-		}
-	}
-
+	D_ASSERT(pool->sp_group != NULL);
 	D_ASSERT(pool->sp_iv_ns != NULL);
+
 	/* Let's invalidate local snapshot cache before
 	 * rebuild, so to make sure rebuild will use the updated
 	 * snapshot during rebuild fetch, otherwise it may cause