This patchset uses bio(s) to read in the journal in large chunks
to lookup the journal head. This will replace the existing
bisect method of reading the journal block-by-block to find the
head.
These are RHEL7 patches because it was simpler for me to
test. Upstream versions shouldn't be very dissimilar. I'll port
once I get some feedback on this.
I'm seeing some promising results with some preliminary testing.
I have 5 lock_nolock filesystems saved off using 'gfs2_edit savemeta'
after killing the node mid-IO so as to save unrecovered journals.
All times are in milliseconds. This table below shows the overall
time taken for recovery and the jhead column represents the time
taken to find the jhead alone. Finding the jhead is a big chunk of
the overall time taken for recovery.
bisect | bio
overall jhead | overall jhead
------------------------------|-----------------------
fs1 1702 1642 | 394 325
fs2 761 334 | 706 356
fs3 1150 904 | 1356 1089
fs4 953 596 | 691 343
fs5 1262 1233 | 554 524
Abhi Das (3):
gfs2: allow map_journal_extents() to take a journal descriptor as
argument
gfs2: add timing info for various stages of journal recovery
gfs2: read journal in large chunks to locate the head
fs/gfs2/incore.h | 7 ++-
fs/gfs2/log.c | 2 +-
fs/gfs2/log.h | 2 +
fs/gfs2/lops.c | 115 +++++++++++++++++++++++++++++++++++++++---
fs/gfs2/lops.h | 1 +
fs/gfs2/ops_fstype.c | 10 ++--
fs/gfs2/recovery.c | 138 +++++++++------------------------------------------
fs/gfs2/recovery.h | 1 +
8 files changed, 149 insertions(+), 127 deletions(-)
--
2.4.11
