Commit 95e98467 authored by Jim Harris's avatar Jim Harris Committed by Daniel Verkamp
Browse files

nvme/overhead: relax ordering around spdk_get_ticks 



The Intel SDM clearly states that only an lfence (not
an mfence) is needed to ensure rdtsc is executed before
subsequent instructions.  So change spdk_mb calls
to spdk_rmb to avoid overhead related to the unneeded
part of the fencing operation.  This provides a more
accurate view of the overhead of both the SPDK and
kernel NVMe drivers.

Signed-off-by: default avatarJim Harris <james.r.harris@intel.com>
Change-Id: Idaeb59db241cc1630e438e300559a02c4d59cd49

Reviewed-on: https://review.gerrithub.io/413148


Tested-by: default avatarSPDK Automated Test System <sys_sgsw@intel.com>
Reviewed-by: default avatarShuhei Matsumoto <shuhei.matsumoto.xt@hitachi.com>
Reviewed-by: default avatarBen Walker <benjamin.walker@intel.com>
Reviewed-by: default avatarDaniel Verkamp <daniel.verkamp@intel.com>
parent 4f0c061d

test/nvme/overhead/overhead.c

+5 −4
Original line number Diff line number Diff line
@@ -293,7 +293,7 @@ submit_single_io(void) 
	offset_in_ios = rand_r(&seed) % entry->size_in_ios;



	start = spdk_get_ticks();

	spdk_mb();

	spdk_rmb();

#if HAVE_LIBAIO

	if (entry->type == ENTRY_TYPE_AIO_FILE) {

		rc = aio_submit(g_ns->u.aio.ctx, &g_task->iocb, entry->u.aio.fd, IO_CMD_PREAD, g_task->buf,
@@ -306,7 +306,7 @@ submit_single_io(void) 
					   entry->io_size_blocks, io_complete, g_task, 0);

	}



	spdk_mb();

	spdk_rmb();

	tsc_submit = spdk_get_ticks() - start;

	g_tsc_submit += tsc_submit;

	if (tsc_submit < g_tsc_submit_min) {
@@ -338,7 +338,8 @@ static void 
check_io(void)

{

	uint64_t end, tsc_complete;

	spdk_mb();



	spdk_rmb();

#if HAVE_LIBAIO

	if (g_ns->type == ENTRY_TYPE_AIO_FILE) {

		aio_check_io();
@@ -347,7 +348,7 @@ check_io(void) 
	{

		spdk_nvme_qpair_process_completions(g_ns->u.nvme.qpair, 0);

	}

	spdk_mb();

	spdk_rmb();

	end = spdk_get_ticks();

	if (g_ns->current_queue_depth == 1) {

		/*