Currently every journal IO is issued as REQ_PREFLUSH | REQ_FUA to
guarantee the ordering requirements the journal has w.r.t. metadata
writeback. THe two ordering constraints are:

1. we cannot overwrite metadata in the journal until we guarantee
that the dirty metadata has been written back in place and is
stable.

2. we cannot write back dirty metadata until it has been written to
the journal and guaranteed to be stable (and hence recoverable) in
the journal.

The ordering guarantees of #1 are provided by REQ_PREFLUSH. This
causes the journal IO to issue a cache flush and wait for it to
complete before issuing the write IO to the journal. Hence all
completed metadata IO is guaranteed to be stable before the journal
overwrites the old metadata.

The ordering guarantees of #2 are provided by the REQ_FUA, which
ensures the journal writes do not complete until they are on stable
storage. Hence by the time the last journal IO in a checkpoint
completes, we know that the entire checkpoint is on stable storage
and we can unpin the dirty metadata and allow it to be written back.

This is the mechanism by which ordering was first implemented in XFS
way back in 2002 by commit 95d97c36e5155075ba2eb22b17562cfcc53fcf96
("Add support for drive write cache flushing") in the xfs-archive
tree.

A lot has changed since then, most notably we now use delayed
logging to checkpoint the filesystem to the journal rather than
write each individual transaction to the journal. Cache flushes on
journal IO are necessary when individual transactions are wholly
contained within a single iclog. However, CIL checkpoints are single
transactions that typically span hundreds to thousands of individual
journal writes, and so the requirements for device cache flushing
have changed.

That is, the ordering rules I state above apply to ordering of
atomic transactions recorded in the journal, not to the journal IO
itself. Hence we need to ensure metadata is stable before we start
writing a new transaction to the journal (guarantee #1), and we need
to ensure the entire transaction is stable in the journal before we
start metadata writeback (guarantee #2).

Hence we only need a REQ_PREFLUSH on the journal IO that starts a
new journal transaction to provide #1, and it is not on any other
journal IO done within the context of that journal transaction.

The CIL checkpoint already issues a cache flush before it starts
writing to the log, so we no longer need the iclog IO to issue a
REQ_REFLUSH for us. Hence if XLOG_START_TRANS is passed
to xlog_write(), we no longer need to mark the first iclog in
the log write with REQ_PREFLUSH for this case. As an added bonus,
this ordering mechanism works for both internal and external logs,
meaning we can remove the explicit data device cache flushes from
the iclog write code when using external logs.

Given the new ordering semantics of commit records for the CIL, we
need iclogs containing commit records to issue a REQ_PREFLUSH. We
also require unmount records to do this. Hence for both
XLOG_COMMIT_TRANS and XLOG_UNMOUNT_TRANS xlog_write() calls we need
to mark the first iclog being written with REQ_PREFLUSH.

For both commit records and unmount records, we also want them
immediately on stable storage, so we want to also mark the iclogs
that contain these records to be marked REQ_FUA. That means if a
record is split across multiple iclogs, they are all marked REQ_FUA
and not just the last one so that when the transaction is completed
all the parts of the record are on stable storage.

And for external logs, unmount records need a pre-write data device
cache flush similar to the CIL checkpoint cache pre-flush as the
internal iclog write code does not do this implicitly anymore.

As an optimisation, when the commit record lands in the same iclog
as the journal transaction starts, we don't need to wait for
anything and can simply use REQ_FUA to provide guarantee #2.  This
means that for fsync() heavy workloads, the cache flush behaviour is
completely unchanged and there is no degradation in performance as a
result of optimise the multi-IO transaction case.

The most notable sign that there is less IO latency on my test
machine (nvme SSDs) is that the "noiclogs" rate has dropped
substantially. This metric indicates that the CIL push is blocking
in xlog_get_iclog_space() waiting for iclog IO completion to occur.
With 8 iclogs of 256kB, the rate is appoximately 1 noiclog event to
every 4 iclog writes. IOWs, every 4th call to xlog_get_iclog_space()
is blocking waiting for log IO. With the changes in this patch, this
drops to 1 noiclog event for every 100 iclog writes. Hence it is
clear that log IO is completing much faster than it was previously,
but it is also clear that for large iclog sizes, this isn't the
performance limiting factor on this hardware.

With smaller iclogs (32kB), however, there is a substantial
difference. With the cache flush modifications, the journal is now
running at over 4000 write IOPS, and the journal throughput is
largely identical to the 256kB iclogs and the noiclog event rate
stays low at about 1:50 iclog writes. The existing code tops out at
about 2500 IOPS as the number of cache flushes dominate performance
and latency. The noiclog event rate is about 1:4, and the
performance variance is quite large as the journal throughput can
fall to less than half the peak sustained rate when the cache flush
rate prevents metadata writeback from keeping up and the log runs
out of space and throttles reservations.

As a result:

	logbsize	fsmark create rate	rm -rf
before	32kb		152851+/-5.3e+04	5m28s
patched	32kb		221533+/-1.1e+04	5m24s

before	256kb		220239+/-6.2e+03	4m58s
patched	256kb		228286+/-9.2e+03	5m06s

The rm -rf times are included because I ran them, but the
differences are largely noise. This workload is largely metadata
read IO latency bound and the changes to the journal cache flushing
doesn't really make any noticable difference to behaviour apart from
a reduction in noiclog events from background CIL pushing.
Signed-off-by: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Chandan Babu R <chandanrlinux@gmail.com>
Reviewed-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Allison Henderson <allison.henderson@oracle.com>
Signed-off-by: Darrick J. Wong <djwong@kernel.org>