The hash tables are only ever set up in newly created journal files, in
space that was just allocated via posix_fallocate() and hence reads as
zeroes anyway, so zeroing the items explicitly is unnecessary.
It is also harmful: the sequential write faults of the memzero() make the
kernel's mmap readahead instantiate the table in ever larger folios, up to
the size of the readahead window, on file systems that support large folios
in the page cache (e.g. XFS, and ext4 since kernel 6.16). And since write
faults on shared file mappings dirty the whole folio, each later update of
a single 16 byte hash bucket then has the whole folio written back. Without
the explicit zeroing, the table's pages are only faulted in when buckets
are first accessed, in random order, and hence mostly as order-0 folios.
Measured on Linux 6.18 with a new journal file (SystemMaxFileSize=50M),
appending 15 syslog-like entries with 23 fields each per second for 120s,
while writing back the file's dirty pages every second, as happens on
machines under page cache pressure, where reclaim keeps waking up the
flusher threads. Bytes written to the (loop) device, median and range of
4-5 runs:
before after
ext4 182.8M (178.4M-185.7M) 41.2M (31.6M-41.9M)
xfs 172.7M (161.0M-182.4M) 40.1M (39.7M-44.9M)
This also avoids writing out the whole data hash table once for every new
journal file, i.e. ~2.8% of the maximum file size.
Related: #40262
(cherry picked from commit 6db7cf38d40e8fe6a11392d2f0cc044ed5acf812)
SecretLUL
29646b6f1c