Re: [PATCH v3 3/4] hw/nvme: add iothread support

[Klaus Jensen]

On Aug 27 17:12, Jinhao Fan wrote:
> Add an option "iothread=x" to do emulation in a seperate iothread.
> This improves the performance because QEMU's main loop is responsible
> for a lot of other work while iothread is dedicated to NVMe emulation.
> Moreover, emulating in iothread brings the potential of polling on
> SQ/CQ doorbells, which I will bring up in a following patch.
> 
> Iothread can be enabled by:
> -object iothread,id=nvme0 \
> -device nvme,iothread=nvme0 \
> 
> Performance comparisons (KIOPS):
> 
> QD         1   4  16  64
> QEMU      41 136 242 338
> iothread  53 155 245 309
> 
> Signed-off-by: Jinhao Fan <fanjinhao21s@ict.ac.cn>
> ---
>  hw/nvme/ctrl.c | 67 ++++++++++++++++++++++++++++++++++++++++++++------
>  hw/nvme/ns.c   | 21 +++++++++++++---
>  hw/nvme/nvme.h |  6 ++++-
>  3 files changed, 82 insertions(+), 12 deletions(-)
> 

In hw/nvme/ns.c you need to guard the blk_flush and blk_drain calls with
an aio_context_acquire and aio_context_release pair.

diff --git a/hw/nvme/ns.c b/hw/nvme/ns.c
index eb9141a67b5c..dcf889f6d5ce 100644
--- a/hw/nvme/ns.c
+++ b/hw/nvme/ns.c
@@ -520,12 +520,21 @@ int nvme_ns_setup(NvmeNamespace *ns, AioContext *ctx, 
Error **errp)

 void nvme_ns_drain(NvmeNamespace *ns)
 {
+    AioContext *ctx = blk_get_aio_context(ns->blkconf.blk);
+
+    aio_context_acquire(ctx);
     blk_drain(ns->blkconf.blk);
+    aio_context_release(ctx);
 }

 void nvme_ns_shutdown(NvmeNamespace *ns)
 {
+    AioContext *ctx = blk_get_aio_context(ns->blkconf.blk);
+
+    aio_context_acquire(ctx);
     blk_flush(ns->blkconf.blk);
+    aio_context_release(ctx);
+
     if (ns->params.zoned) {
         nvme_zoned_ns_shutdown(ns);
     }

Otherwise, it all looks fine. I'm still seeing the weird slowdown when
an iothread is enabled. I have yet to figure out why that is... But it
scales! :)

signature.asc
Description: PGP signature