> On Dec 3, 2021, at 4:55 PM, Bruce Fields <bfields@xxxxxxxxxxxx> wrote:
>
> From: "J. Bruce Fields" <bfields@xxxxxxxxxx>
>
> Currently nfsdcld is doing three fdatasyncs for each upcall. Based on
> SQLite documentation, WAL mode should also be safe, and I can confirm
> from an strace that it results in only one fdatasync each.
>
> This may be a bottleneck e.g. when lots of clients are being created or
> expired at once (e.g. on reboot).
>
> Not bothering with error checking, as this is just an optimization and
> nfsdcld will still function without. (Might be better to log something
> on failure, though.)
I'm in full philosophical agreement for performance improvements
in this area. There are some caveats for WAL:
- It requires SQLite v3.7.0 (2010). configure.ac may need to be
updated.
- All processes using the DB file have to be on the same system.
Not sure if this matters for some DR scenarios that nfs-utils
is supposed to support.
- WAL mode is persistent; you could set this at database creation
time and it should be sticky.
- Documentation says "synchronous = FULL is the most commonly
used synchronous setting when not in WAL mode." Why are both
PRAGMAs necessary here?
I agree that nfsdcld functionality is not affected if the first
PRAGMA fails.
> Signed-off-by: J. Bruce Fields <bfields@xxxxxxxxxx>
> ---
> utils/nfsdcld/sqlite.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> diff --git a/utils/nfsdcld/sqlite.c b/utils/nfsdcld/sqlite.c
> index 03016fb95823..b248eeffa204 100644
> --- a/utils/nfsdcld/sqlite.c
> +++ b/utils/nfsdcld/sqlite.c
> @@ -826,6 +826,9 @@ sqlite_prepare_dbh(const char *topdir)
> goto out_close;
> }
>
> + sqlite3_exec(dbh, "PRAGMA journal_mode = WAL;", NULL, NULL, NULL);
> + sqlite3_exec(dbh, "PRAGMA synchronous = FULL;", NULL, NULL, NULL);
> +
> ret = sqlite_query_schema_version();
> switch (ret) {
> case CLD_SQLITE_LATEST_SCHEMA_VERSION:
> --
> 2.33.1
>
--
Chuck Lever
