InnoDB: Failing assertion: mode == BUF_GET_POSSIBLY_FREED || !block->page.file_page_was_freed in file buf0buf.c line 2910

Valgrind in buf_page_get_gen in bug 1086680

lock_print_info_all_transactions attempts to load a buffer page for a possible eventual detailed lock dump:

   lock_mutex_exit_kernel();

   mtr_start(&mtr);

   buf_page_get_with_no_latch(space, zip_size,
         page_no, &mtr);

   mtr_commit(&mtr);

   load_page_first = FALSE;

   lock_mutex_enter_kernel();

First of all, this does not seem to be necessary at all when XtraDB-specific innodb_show_verbose_locks is 0 (which it is in this case).
Second, the buffer page read happens with no locks held, thus the read must handle everything, including a disappearing page, which is the cause of this crash. The latter seems to apply to InnoDB too.

The redundant page reads for innodb_show_verbose_locks=0 is bug 1100643 now.

Confirmed for 5.1 by code analysis.

The upstream fixed other instances this assertion might come up by the following commit in 5.5.

$ bzr log -r 3400
------------------------------------------------------------
revno: 3400
committer: Marko Mäkelä <email address hidden>
branch nick: mysql-5.5
timestamp: Tue 2011-05-24 23:08:22 +0300
message:
  Bug#12584374 LOCK_VALIDATE TRIPS ASSERTION !BLOCK->PAGE.FILE_PAGE_WAS_FREE

  Fix a deadlock in the initial patch. lock_validate() must not hold the
  lock system mutex while s-latching a block, because some functions,
  such as lock_rec_convert_impl_to_expl(), may be already holding an x-latch
  on the block that lock_validate() is interested in while attempting to
  acquire the lock system mutex.

  This deadlock was not caught by UNIV_SYNC_DEBUG because of
  buf_block_dbg_add_level(block, SYNC_NO_ORDER_CHECK).

I cannot find any attempts to fix this for 5.1 in the upstream.

Roel has testcase (same one as for bug bug1086680 / bug 1083058).

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PS-616