[Qemu-devel] [PATCH v2 0/2] virtio-scsi: Optimizing request allocation

[Qemu-devel] [PATCH v2 0/2] virtio-scsi: Optimizing request allocation

[Fam Zheng]

v2: Slight improvements according to Paolo's comments.
    4k -> 48k.

Zeroing is relatively expensive since we have big request structures.
VirtQueueElement (>48k!) and sense_buf (256 bytes) are two points to look at.

This visibly reduces overhead of request handling when testing with the
unmerged "null" driver and virtio-scsi dataplane. Before, the issue is very
obvious with perf top:

perf top -G -p `pidof qemu-system-x86_64`
-----------------------------------------
+  16.50%  libc-2.17.so             [.] __memset_sse2
+   2.28%  libc-2.17.so             [.] _int_malloc
+   2.25%  [vdso]                   [.] 0x0000000000000cd1
+   2.02%  [kernel]                 [k] _raw_spin_lock_irqsave
+   1.97%  libpthread-2.17.so       [.] pthread_mutex_lock
+   1.87%  libpthread-2.17.so       [.] pthread_mutex_unlock
+   1.81%  [kernel]                 [k] fget_light
+   1.70%  libc-2.17.so             [.] malloc

After, the high __memset_sse2 and _int_malloc is gone:

perf top -G -p `pidof qemu-system-x86_64`
-----------------------------------------
+   4.20%  [kernel]                 [k] vcpu_enter_guest
+   3.97%  [kernel]                 [k] vmx_vcpu_run
+   2.63%  [kernel]                 [k] _raw_spin_lock_irqsave
+   1.72%  [kernel]                 [k] native_read_msr_safe
+   1.65%  [kernel]                 [k] __srcu_read_lock
+   1.64%  [kernel]                 [k] _raw_spin_unlock_irqrestore
+   1.57%  [vdso]                   [.] 0x00000000000008d8
+   1.49%  libc-2.17.so             [.] _int_malloc
+   1.29%  libpthread-2.17.so       [.] pthread_mutex_unlock
+   1.26%  [kernel]                 [k] native_write_msr_safe

See the commit message of patch 2 for some fio test data.

Thanks,
Fam

Fam Zheng (2):
  scsi: Optimize scsi_req_alloc
  virtio-scsi: Optimize virtio_scsi_init_req

 hw/scsi/scsi-bus.c     |  8 +++++---
 hw/scsi/virtio-scsi.c  | 23 +++++++++++++++++------
 include/hw/scsi/scsi.h | 21 ++++++++++++++-------
 3 files changed, 36 insertions(+), 16 deletions(-)

-- 
1.9.3

[Qemu-devel] [PATCH v2 1/2] scsi: Optimize scsi_req_alloc

[Fam Zheng]

Zeroing sense buffer for each scsi request is not efficient, we can just
leave it uninitialized because sense_len is set to 0.

Move the implicitly zeroed fields to the end of the structure and use a
partial memset.

The explicitly initialized fields (by scsi_req_alloc or scsi_req_new)
are moved to the beginning of the structure, before sense buffer, to
skip the memset.

Also change g_malloc0 to g_slice_alloc.

Signed-off-by: Fam Zheng <famz@redhat.com>
---
 hw/scsi/scsi-bus.c     |  8 +++++---
 include/hw/scsi/scsi.h | 21 ++++++++++++++-------
 2 files changed, 19 insertions(+), 10 deletions(-)

diff --git a/hw/scsi/scsi-bus.c b/hw/scsi/scsi-bus.c
index 954c607..af293b5 100644
--- a/hw/scsi/scsi-bus.c
+++ b/hw/scsi/scsi-bus.c
@@ -551,8 +551,11 @@ SCSIRequest *scsi_req_alloc(const SCSIReqOps *reqops, SCSIDevice *d,
     SCSIRequest *req;
     SCSIBus *bus = scsi_bus_from_device(d);
     BusState *qbus = BUS(bus);
+    const int memset_off = offsetof(SCSIRequest, sense)
+                           + sizeof(req->sense);
 
-    req = g_malloc0(reqops->size);
+    req = g_slice_alloc(reqops->size);
+    memset((uint8_t *)req + memset_off, 0, reqops->size - memset_off);
     req->refcount = 1;
     req->bus = bus;
     req->dev = d;
@@ -560,7 +563,6 @@ SCSIRequest *scsi_req_alloc(const SCSIReqOps *reqops, SCSIDevice *d,
     req->lun = lun;
     req->hba_private = hba_private;
     req->status = -1;
-    req->sense_len = 0;
     req->ops = reqops;
     object_ref(OBJECT(d));
     object_ref(OBJECT(qbus->parent));
@@ -1603,7 +1605,7 @@ void scsi_req_unref(SCSIRequest *req)
         }
         object_unref(OBJECT(req->dev));
         object_unref(OBJECT(qbus->parent));
-        g_free(req);
+        g_slice_free1(req->ops->size, req);
     }
 }
 
diff --git a/include/hw/scsi/scsi.h b/include/hw/scsi/scsi.h
index 2e3a8f9..2a7d4e5 100644
--- a/include/hw/scsi/scsi.h
+++ b/include/hw/scsi/scsi.h
@@ -50,17 +50,24 @@ struct SCSIRequest {
     uint32_t          tag;
     uint32_t          lun;
     uint32_t          status;
+    void              *hba_private;
     size_t            resid;
+
+    /* Note:
+     * - fields before sense are initialized by scsi_req_alloc;
+     * - sense[] is uninitialized;
+     * - fields after sense are memset to 0 by scsi_req_alloc.
+     * */
+
+    uint8_t           sense[SCSI_SENSE_BUF_SIZE];
+    uint32_t          sense_len;
+    bool              enqueued;
+    bool              io_canceled;
+    bool              retry;
+    bool              dma_started;
     SCSICommand       cmd;
     BlockDriverAIOCB  *aiocb;
     QEMUSGList        *sg;
-    bool              dma_started;
-    uint8_t sense[SCSI_SENSE_BUF_SIZE];
-    uint32_t sense_len;
-    bool enqueued;
-    bool io_canceled;
-    bool retry;
-    void *hba_private;
     QTAILQ_ENTRY(SCSIRequest) next;
 };
 
-- 
1.9.3

Re: [Qemu-devel] [PATCH v2 1/2] scsi: Optimize scsi_req_alloc

[Paolo Bonzini]

Il 15/09/2014 07:23, Fam Zheng ha scritto:
> +    uint8_t           sense[SCSI_SENSE_BUF_SIZE];
> +    uint32_t          sense_len;
> +    bool              enqueued;
> +    bool              io_canceled;
> +    bool              retry;
> +    bool              dma_started;
>      SCSICommand       cmd;
>      BlockDriverAIOCB  *aiocb;
>      QEMUSGList        *sg;
> -    bool              dma_started;
> -    uint8_t sense[SCSI_SENSE_BUF_SIZE];
> -    uint32_t sense_len;
> -    bool enqueued;
> -    bool io_canceled;
> -    bool retry;
> -    void *hba_private;
>      QTAILQ_ENTRY(SCSIRequest) next;

Similar to v1, cmd is filled in by scsi_req_new and can be moved to the
uninitialized area.

Paolo

[Qemu-devel] [PATCH v2 2/2] virtio-scsi: Optimize virtio_scsi_init_req

[Fam Zheng]

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 | 23 +++++++++++++++++------
 1 file changed, 17 insertions(+), 6 deletions(-)

diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c
index 86aba88..7bf03c4 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 +75,27 @@ 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

Re: [Qemu-devel] [PATCH v2 2/2] virtio-scsi: Optimize virtio_scsi_init_req

[Paolo Bonzini]

Il 15/09/2014 07:23, Fam Zheng ha scritto:
>      SCSIRequest *sreq;
>      size_t resp_size;
>      enum SCSIXferMode mode;
> -    QEMUIOVector resp_iov;
>      union {
>          VirtIOSCSICmdResp     cmd;
>          VirtIOSCSICtrlTMFResp tmf;
> @@ -68,23 +75,27 @@ 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);

Looks good, but why do you need to zero the union?  You only need to
zero sreq, resp_size and mode, don't you (and at this point, memset
becomes superfluous)?

Paolo

Re: [Qemu-devel] [PATCH v2 2/2] virtio-scsi: Optimize virtio_scsi_init_req

[Fam Zheng]

On Mon, 09/15 12:17, Paolo Bonzini wrote:
> Il 15/09/2014 07:23, Fam Zheng ha scritto:
> >      SCSIRequest *sreq;
> >      size_t resp_size;
> >      enum SCSIXferMode mode;
> > -    QEMUIOVector resp_iov;
> >      union {
> >          VirtIOSCSICmdResp     cmd;
> >          VirtIOSCSICtrlTMFResp tmf;
> > @@ -68,23 +75,27 @@ 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);
> 
> Looks good, but why do you need to zero the union?  You only need to
> zero sreq, resp_size and mode, don't you (and at this point, memset
> becomes superfluous)?
> 

The structures in unions are not zeroed by caller, also leaving them breaks
virtio-scsi in my test.

FWIW, I will remove the "req->sreq = NULL;" two lines below in v3. At this
point tuning these small fields are subtle optimization compared to the arrays,
I say let's just simply keep the memset so that adding more fields in the
future are also safe.

Fam

Re: [Qemu-devel] [PATCH v2 2/2] virtio-scsi: Optimize virtio_scsi_init_req

[Paolo Bonzini]

Il 16/09/2014 09:16, Fam Zheng ha scritto:
> On Mon, 09/15 12:17, Paolo Bonzini wrote:
>> Il 15/09/2014 07:23, Fam Zheng ha scritto:
>>>      SCSIRequest *sreq;
>>>      size_t resp_size;
>>>      enum SCSIXferMode mode;
>>> -    QEMUIOVector resp_iov;
>>>      union {
>>>          VirtIOSCSICmdResp     cmd;
>>>          VirtIOSCSICtrlTMFResp tmf;
>>> @@ -68,23 +75,27 @@ 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);
>>
>> Looks good, but why do you need to zero the union?  You only need to
>> zero sreq, resp_size and mode, don't you (and at this point, memset
>> becomes superfluous)?
>>
> 
> The structures in unions are not zeroed by caller, also leaving them breaks
> virtio-scsi in my test.
> 
> FWIW, I will remove the "req->sreq = NULL;" two lines below in v3. At this
> point tuning these small fields are subtle optimization compared to the arrays,
> I say let's just simply keep the memset so that adding more fields in the
> future are also safe.

Perhaps the response fields have to be zeroed?  The request shouldn't
need it.  It can be done separately though---the VirtQueueElement is the
big one that we have to fix.

Paolo