Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-20816

osd-ldiskfs: faked read unlocks a cached DIO folio

XMLWordPrintable

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

      A read faked with OBD_FAIL_OST_FAKE_RW (0x238) marks its folio
      uptodate, so osd_read_prep() takes the cache-hit branch and unlocks
      the folio.

      The per-thread DIO folios are locked when allocated and stay locked
      for as long as the thread caches them; osd_get_folio() reuses them
      without locking again. Unlocking one in that branch leaves it
      unlocked for its next use, and the next user asserts:

      ASSERTION( folio_test_locked(folio) ) failed

      Three sites can hit it, all in master:

      osd_write_commit() lustre/osd-ldiskfs/osd_io.c:1523, 1563
      osd_read_prep() lustre/osd-ldiskfs/osd_io.c:1652
      osd_key_fini() lustre/osd-ldiskfs/osd_handler.c:8290

      so a real write or a cache-miss read reusing the folio LBUGs, and
      failing that, the OST LBUGs on unmount when the thread key is torn
      down.

      Reproduced on a single node OST with fail_loc=0x238 set and direct
      I/O reads and writes running against it.

      A real read cannot reach the cache-hit branch holding a DIO folio,
      since osd_get_folio() clears uptodate on each reuse, and page cache
      folios never carry private_2. Only the faked read gets there.

            patrickbot Patrick Bot
            patrickbot Patrick Bot
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

              Created:
              Updated: