[LU-10859] Deadlock with heavy memory pressure

Related — Mon, 11 May 2020 17:56:04 +0000

On one Customer site, we hit following deadlock:

    Thread 1:

    ofd_object_punch

     osd_punch

      ldiskfs_truncate

       ldiskfs_inode_attach_jinode

         ...

         do_try_to_free_pages

          lu_cache_shrink

           mutex_lock -->try to hold @lu_sites_guard

    

    kswapd thread2:

    kthread

     shrink_slab

      lu_cache_shrink

        mutex_lock ---->hold already.

         ...

         dqget

          ldiskfs_acquire_dquot

           jbd2__journal_start-->blocked to wait for more credits.

    

    Thread3:

    kthread

     kjournald2

      jbd2_journal_commit_transaction-->blocked to wait Thread2 finished,

                                 since Thread1 add a handle into transaction.

    

    So deadlock happens because of Thread1 wait Thread2, Thread2 wait Thread3..

    but Thread3 wait Thread1....

    

    This problem still exists even we have switched @lu_sites_guard

    into a read/write lock, sine we hold write lock at lu_cahce_shrink().

    

    Fixed the problem by making ldiskfs_inode_attach_jinode() use

    GPF_NOFS.

Whamcloud Community JIRA

[LU-10859] Deadlock with heavy memory pressure