Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-21266

o2iblnd: NULL list_del in kiblnd_destroy_tx_pool after ENOMEM

XMLWordPrintable

    • Icon: Bug Bug
    • Resolution: Unresolved
    • Icon: Medium Medium
    • None
    • None
    • None
    • 3
    • 9223372036854775807

      An allocation failure in kiblnd_create_tx_pool() makes kiblnd_destroy_tx_pool() dereference a NULL pointer.

      Mechanism

      kiblnd_create_tx_pool() zeroes all tpo_tx_descs with memset(). It then allocates the per-tx arrays in a loop. On an allocation failure, the loop stops with break, and the function calls ps->ps_pool_destroy().

      kiblnd_destroy_tx_pool() calls list_del(&tx->tx_list) for each of the po_size descriptors. Only kiblnd_map_tx_pool() adds a descriptor to po_free_list, and the create function calls it only after the loop completes. On the failure path, every tx_list is still zero. list_del() then writes through a NULL next pointer.

      kiblnd_create_tx_pool()
        memset(tpo->tpo_tx_descs, 0, ...)
        for (i = 0; i < size; i++)
            LIBCFS_CPT_ALLOC(...) fails -> break
        ps->ps_pool_destroy(pool)
          kiblnd_destroy_tx_pool()
            list_del(&tx->tx_list)   <- tx_list.next == NULL
      

      Trigger

      Any of the per-tx allocations (tx_pages, tx_frags, tx_wrq, tx_sge, tx_rd) fails under memory pressure. The failure can happen at LNet startup or when the pool set grows.

      Fix

      Initialize tx_list of each descriptor before the allocation loop, so that list_del() is safe on the failure path.

      The defect is source-only. Nobody has seen the crash in a log. A code review found the defect on current master.

            hornc Chris Horn
            hornc Chris Horn
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

              Created:
              Updated: