Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-21075

llite: vvp_prep_size() shrinks i_size down to KMS, a racing stat() reports a PFL component boundary

XMLWordPrintable

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

      vvp_prep_size() writes KMS into i_size unconditionally:

      i_size_write(inode, kms);
      KMS is a lower bound only. osc_lock_lvb_update() clamps it to the end of
      the granted DLM extent, and sub-objects with no valid KMS contribute nothing
      to the merged value. On a PFL file whose size lives in a component the client
      holds no lock on, that publishes the end of the component being read as the
      file size.

      A concurrent stat() reads i_size back out of the inode and reports it.
      A 4K read at offset 0 of a 200MB file whose first component ends at 128K
      leaves i_size at 131072, while the cat_size of the very same
      cl_attr says 209715200.

      Where it comes from
      The unconditional write was introduced by:

      ec07ddc4120b9850982c8f39bcc43e57d2d76549
      LUS-13048 llite: set read size correctly
      Before it, i_size was only ever extended:

      if (i_size_read(inode) < kms)
      i_size_write(inode, kms);
      LUS-13048 made the write unconditional so that a size which had decreased
      after a lock renewal would be followed - sanityn test_16i, a read after a
      truncate done by another client. That part has to be kept; what must go is
      lowering i_size to a value that is not a size at all.

      Fix
      Extend i_size up to KMS, and use cat_size as the upper bound rather
      than KMS:

      cat_size is max(lvb_size, loi_kms) per stripe, merged as a max over
      the instantiated components, so it is always >= KMS;
      the OSC layer lowers it only on an authoritative update, such as another
      client's truncate.
      Keeping that clamp is what stops the generic read loop - which runs to
      i_size with no further control from llite - from returning zeroes between
      a new EOF and a stale i_size after a remote truncate, i.e. sanityn
      test_16i still passes.

      i_size is sampled once for both comparisons, since ll_update_inode()
      writes it without ll_inode_size_lock() and a second read may differ; the
      debug lines print the value actually written.

            shadow Alexey Lyashkov
            shadow Alexey Lyashkov
            Votes:
            0 Vote for this issue
            Watchers:
            4 Start watching this issue

              Created:
              Updated: