Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-20679

Possible long processing of OBD_IDX_READ RPC

XMLWordPrintable

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

      The below is description of a real case:

      The LFSCK orphan scan on OST002b was initiated by MDT0002, however the vast majority of objects on that OST (~163M objects across 5,302 rbtree nodes) belong to sequence 0xe80000402 which is owned by MDT0001. As a result, lfsck_orphan_it_next() was performing expensive per-object disk lookups (ldiskfs directory search + LMA xattr + PFID xattr reads) for every single object, only to discard each one because it belongs to a different MDT. No objects were being added to the reply buffer, causing the OBD_IDX_READ RPC to spin for 40+ minutes without making any forward progress.
      

      It would make sense to change tgt_obd_idx_read() so that it respected the RPC deadline and return a partial reply (even if the buffer is empty) along with the last scanned position.

            wc-triage WC Triage
            vsaveliev Vladimir Saveliev
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

              Created:
              Updated: