commit d7eb1eeff0c029ebf7ffcf8b6cc35dd87f62e60c
author Chris Suter <csuter@fuchsia.infra.roller.google.com> Tue Oct 12 18:26:34 2021 +0000
committer Copybara-Service <copybara-worker@google.com> Tue Oct 12 11:27:55 2021 -0700
tree 1a88146e5618a428d5b3fb0b637ebeda0d09f438
parent 97a12e28a4b8574bf8c5aaf30613b865896cb7f3

[roll] Roll fuchsia [fxfs] Fix and enable fxfs-directory-large-tests

This CL includes a number of changes to get some of the slow tests
running:

 + Added a test for a shrink race that should be fixed now.

 + Optimised DirectWriter so that it buffers much more data. This
   substantially reduces the amount of metadata required for writing
   layer files.

 + Fixed a bug with updating reservations after compaction. Prior to
   this CL, it was updating the reservations required with the *old*
   layer sizes rather than the new ones. Fixing this involved adding
   an associated object that contains the reservation update.

 + Changed the journal so that after replaying it guarantees that all
   stores are open, which means we no longer need to worry about calling
   ensure_open in so many places and it doesn't need to be thread-safe.

 + Fixed some checks in the allocator that weren't checking all the
   right values. Added a new taken_bytes() function to reduce the chance
   of getting this wrong in future.

 + Added context to all inconsistent errors.

 + Fixed a small reservation inconsistent between unmount and mount.
   After remount of a cleanly shut down filesystem, the padding in the
   journal would get borrowed, but that's not what would have happened
   if there'd been no unmount. This is fixed by storing the offset
   prior to padding in the super-block.

 + reserved_space_from_journal_usage was not accounting for the
   additional space required for compactions.
   reserved_space_from_journal_usage now returns four times the amount
   used in the journal.

 + The amount of space to reserve for compactions did not include any
   space required for metadata. A new function,
   reservation_amount_from_layer_size, adds a 50% overhead.

 + This CL changes reads to use streams which is required to avoid the
   shrink race mentioned above.

Original-Reviewed-on: https://fuchsia-review.googlesource.com/c/fuchsia/+/592961
Original-Revision: 5ef6bc1b02d3b95c4ca7f77256a7cf823223a953
GitOrigin-RevId: be64b4fa47a10df1d74329d0de76536682db50bb
Change-Id: Ide6e568684da0b2fdcf7cb3e767cc14f219cec59