[Qemu-devel] [PATCH v3 2/2] virtio-scsi: Optimize virtio_scsi_init_req

[Fam Zheng <famz@redhat.com>]

The VirtQueueElement is a very big structure (>48k!), since it will be
initialzed by virtqueue_pop, we can save the expensive zeroing here.

This saves a few microseconds per request in my test:

[fio-test]      rw         bs         iodepth    jobs       bw         iops       latency
--------------------------------------------------------------------------------------------
Before          read       4k         1          1          110        28269      34
After           read       4k         1          1          131        33745      28

Whereas,

virtio-blk      read       4k         1          1          217        55673      16

Signed-off-by: Fam Zheng <famz@redhat.com>
---
 hw/scsi/virtio-scsi.c | 24 +++++++++++++++++-------
 1 file changed, 17 insertions(+), 7 deletions(-)

diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c
index 86aba88..f0d21a3 100644
--- a/hw/scsi/virtio-scsi.c
+++ b/hw/scsi/virtio-scsi.c
@@ -24,12 +24,19 @@
 typedef struct VirtIOSCSIReq {
     VirtIOSCSI *dev;
     VirtQueue *vq;
-    VirtQueueElement elem;
     QEMUSGList qsgl;
+    QEMUIOVector resp_iov;
+
+    /* Note:
+     * - fields before elem are initialized by virtio_scsi_init_req;
+     * - elem is uninitialized at the time of allocation.
+     * - fields after elem are zeroed by virtio_scsi_init_req.
+     * */
+
+    VirtQueueElement elem;
     SCSIRequest *sreq;
     size_t resp_size;
     enum SCSIXferMode mode;
-    QEMUIOVector resp_iov;
     union {
         VirtIOSCSICmdResp     cmd;
         VirtIOSCSICtrlTMFResp tmf;
@@ -68,23 +75,26 @@ static inline SCSIDevice *virtio_scsi_device_find(VirtIOSCSI *s, uint8_t *lun)
 static VirtIOSCSIReq *virtio_scsi_init_req(VirtIOSCSI *s, VirtQueue *vq)
 {
     VirtIOSCSIReq *req;
-    VirtIOSCSICommon *vs = VIRTIO_SCSI_COMMON(s);
-
-    req = g_malloc0(sizeof(*req) + vs->cdb_size);
+    VirtIOSCSICommon *vs = (VirtIOSCSICommon *)s;
+    const size_t zero_skip = offsetof(VirtIOSCSIReq, elem)
+                             + sizeof(VirtQueueElement);
 
+    req = g_slice_alloc(sizeof(*req) + vs->cdb_size);
     req->vq = vq;
     req->dev = s;
-    req->sreq = NULL;
     qemu_sglist_init(&req->qsgl, DEVICE(s), 8, &address_space_memory);
     qemu_iovec_init(&req->resp_iov, 1);
+    memset((uint8_t *)req + zero_skip, 0, sizeof(*req) - zero_skip);
     return req;
 }
 
 static void virtio_scsi_free_req(VirtIOSCSIReq *req)
 {
+    VirtIOSCSICommon *vs = (VirtIOSCSICommon *)req->dev;
+
     qemu_iovec_destroy(&req->resp_iov);
     qemu_sglist_destroy(&req->qsgl);
-    g_free(req);
+    g_slice_free1(sizeof(*req) + vs->cdb_size, req);
 }
 
 static void virtio_scsi_complete_req(VirtIOSCSIReq *req)
-- 
1.9.3