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

[Paolo Bonzini <pbonzini@redhat.com>]

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