Re: [Qemu-devel] [RFC PATCH v1 0/5] Enable virtio-scsi boot from /dev/sgX

[Christian Borntraeger <borntraeger@de.ibm.com>]

On 05/06/2017 10:24 AM, Paolo Bonzini wrote:
> 
> 
> On 05/05/2017 18:12, Eric Farman wrote:
>>
>>
>> On 05/05/2017 11:13 AM, Paolo Bonzini wrote:
>>>
>>>
>>> On 05/05/2017 17:03, Eric Farman wrote:
>>>> We get a value of x3fffff when sending that to a scsi-disk from bios
>>>> code.  That's fully emulated though, in scsi_disk_emulate_inquiry.  And
>>>> that's the scenario that already works.
>>>>
>>>> While there is indeed code in hw/scsi/scsi-generic.c to wire that in,
>>>> that only happens after the I/O goes to the device itself.  The Block
>>>> Limits page isn't supported [1] and thus it gets rejected with "invalid
>>>> field in cdb".  We never get to that fixup code you reference, since the
>>>> returned len is zero.
>>>>
>>>> Should I be refactoring this code to always patch in that block limit
>>>> regardless of a response from the host/device?  (That is, when page xb0
>>>> isn't supported by the hw.)
>>>
>>> What is thec value you get?
>>
>> x140000 bytes when using /dev/sg0 (xa00 sectors when using /dev/sda).
>>
>>>  Is there a sensible default value
>>> that you can use when page 0xb0 isn't supported by the hardware?
>>
>> I was setting max_sectors to x800 with good success, which was the
>> power-of-2 floor that BLKSECTGET gave us.  That kept us within the
>> limits of the host biovec code.  But it's a long way from the
>> virtio-scsi value of xFFFF when max_sectors isn't specified, so don't
>> know what side effects that may cause.
> 
> It's just slower, but 0x800 is already a megabyte worth of data so it's
> not going to be that much slower.
> 
> Paolo
> 

Eric, maybe that would be a safe variant. The bios boot code is not
performance optimized anyway. Lets just always use a maximum size that 
will always work.