MDEV-24391 heap-use-after-free in fil_space_t::flush_low()

We observed a race condition that involved two threads
executing fil_flush_file_spaces() and one thread
executing fil_delete_tablespace(). After one of the
fil_flush_file_spaces() observed that
space.needs_flush_not_stopping() is set and was
releasing the fil_system.mutex, the other fil_flush_file_spaces()
would complete the execution of fil_space_t::flush_low() on
the same tablespace. Then, fil_delete_tablespace() would
destroy the object, because the value of fil_space_t::n_pending
did not prevent that. Finally, the fil_flush_file_spaces() would
resume execution and invoke fil_space_t::flush_low() on the freed
object.

This race condition was introduced in
commit 118e258aaa of MDEV-23855.

fil_space_t::flush(): Add a template parameter that indicates
whether the caller is holding a reference to prevent the
tablespace from being freed.

buf_dblwr_t::flush_buffered_writes_completed(),
row_quiesce_table_start(): Acquire a reference for the duration
of the fil_space_t::flush_low() operation. It should be impossible
for the object to be freed in these code paths, but we want to
satisfy the debug assertions.

fil_space_t::flush_low(): Do not increment or decrement the
reference count, but instead assert that the caller is holding
a reference.

fil_space_extend_must_retry(), fil_flush_file_spaces():
Acquire a reference before releasing fil_system.mutex.
This is what will fix the race condition.

extra/mariabackup/xtrabackup.cc

@@ -552,7 +552,7 @@ void CorruptedPages::zero_out_free_pages()
             *page_it, space_name.c_str());
       }
     }
-    space->flush();
+    space->flush<true>();
     space->release();
   }
   m_spaces.swap(non_free_pages);

storage/innobase/buf/buf0dblwr.cc

@@ -648,7 +648,7 @@ void buf_dblwr_t::flush_buffered_writes_completed(const IORequest &request)
   mysql_mutex_unlock(&mutex);
 
   /* Now flush the doublewrite buffer data to disk */
-  fil_system.sys_space->flush();
+  fil_system.sys_space->flush<false>();
 
   /* The writes have been flushed to disk now and in recovery we will
   find them in the doublewrite buffer blocks. Next, write the data pages. */

storage/innobase/fil/fil0fil.cc

@@ -496,18 +496,16 @@ void fil_space_t::flush_low()
 {
   ut_ad(!mutex_own(&fil_system.mutex));
 
-  uint32_t n= 0;
-  while (!n_pending.compare_exchange_strong(n, (n + 1) | NEEDS_FSYNC,
+  uint32_t n= 1;
+  while (!n_pending.compare_exchange_strong(n, n | NEEDS_FSYNC,
                                             std::memory_order_acquire,
                                             std::memory_order_relaxed))
   {
+    ut_ad(n & PENDING);
     if (n & STOPPING)
       return;
-    if (!(n & NEEDS_FSYNC))
-      continue;
-    if (acquire_low() & STOPPING)
-      return;
-    break;
+    if (n & NEEDS_FSYNC)
+      break;
   }
 
   fil_n_pending_tablespace_flushes++;
@@ -535,7 +533,6 @@ void fil_space_t::flush_low()
   }
 
   clear_flush();
-  release();
   fil_n_pending_tablespace_flushes--;
 }
 
@@ -632,8 +629,10 @@ fil_space_extend_must_retry(
 	case TRX_SYS_SPACE:
 		srv_sys_space.set_last_file_size(pages_in_MiB);
 	do_flush:
+		space->reacquire();
 		mutex_exit(&fil_system.mutex);
 		space->flush_low();
+		space->release();
 		mutex_enter(&fil_system.mutex);
 		break;
 	default:
@@ -3516,8 +3515,10 @@ rescan:
   {
     if (space.needs_flush_not_stopping())
     {
+      space.reacquire();
       mutex_exit(&fil_system.mutex);
       space.flush_low();
+      space.release();
       goto rescan;
     }
   }

storage/innobase/include/fil0fil.h

@@ -972,7 +972,7 @@ public:
   fil_io_t io(const IORequest &type, os_offset_t offset, size_t len,
               void *buf, buf_page_t *bpage= nullptr);
   /** Flush pending writes from the file system cache to the file. */
-  inline void flush();
+  template<bool have_reference> inline void flush();
   /** Flush pending writes from the file system cache to the file. */
   void flush_low();
 
@@ -1452,18 +1452,23 @@ inline void fil_space_t::set_stopping(bool stopping)
 }
 
 /** Flush pending writes from the file system cache to the file. */
-inline void fil_space_t::flush()
+template<bool have_reference> inline void fil_space_t::flush()
 {
   ut_ad(!mutex_own(&fil_system.mutex));
-
+  ut_ad(!have_reference || (pending() & PENDING));
   ut_ad(purpose == FIL_TYPE_TABLESPACE || purpose == FIL_TYPE_IMPORT);
   if (srv_file_flush_method == SRV_O_DIRECT_NO_FSYNC)
   {
     ut_ad(!is_in_unflushed_spaces);
     ut_ad(!needs_flush());
   }
-  else
+  else if (have_reference)
+    flush_low();
+  else if (!(acquire_low() & STOPPING))
+  {
     flush_low();
+    release();
+  }
 }
 
 /** @return the size in pages (0 if unreadable) */

storage/innobase/row/row0quiesce.cc

@@ -545,7 +545,7 @@ row_quiesce_table_start(
 	if (!trx_is_interrupted(trx)) {
 		/* Ensure that all asynchronous IO is completed. */
 		os_aio_wait_until_no_pending_writes();
-		table->space->flush();
+		table->space->flush<false>();
 
 		if (row_quiesce_write_cfg(table, trx->mysql_thd)
 		    != DB_SUCCESS) {