Re: [Qemu-devel] [PATCH] Pass the drive's readonly attribute to the guest OS

[Kevin Wolf]

Am 15.10.2009 12:11, schrieb Gleb Natapov:
> On Thu, Oct 15, 2009 at 12:05:41PM +0200, Kevin Wolf wrote:
>>>>>>>> If the right response to a write on a read-only device is defined in 
>>>>>>>> the
>>>>>>>> specification (and it most probably is), we should still give the right
>>>>>>>> response, even though the OS is doing something wrong.
>>>>>>>>
>>>>>>> And since our response to write error may be pausing a VM we shouldn't
>>>>>>> allow this to be triggered by a guest OS.
>>>>>>
>>>>>> I thought we only pause the VM if we get an host IO error? But if you do
>>>>>> want to stop it for all errors, you shouldn't start suppressing errors
>>>>>> so that it doesn't stop.
>>>>>>
>>>>> We pause only on host IO errors, but if we open underlying file as
>>>>> read only (do we?) and try to write into it we will get an IO error
>>>>> in the host.
>>>>
>>>> No, we'll return an error before a write request to the host is even 
>>>> issued.
>>>>
>>> Who is "we"? If "we" == "bdrv_write()/dma_bdrv_write()" then it's all the 
>>> same.
>>> Upper layers don't actually care why block driver failed.
>>
>> Right, "we" is the qemu block layer. If the devices don't use the error
>> code returned, they better should be fixed, I think?
>>
> Fixed in what way? There is an option to pause VM on any write error. If
> block layer returns write error for whatever reason VM will be paused.
> If scsi/ide/virtio knows that the media is read only it shouldn't
> issue writes to block layer, but handle it like real HW would.

Ok, I agree if you want to do it this way. How it's done I really don't
care too much, but in the end the guest OS should see the right error. I
had the impression that Naphtali and you wanted to silently ignore
writes to a read-only disk, which would be just wrong.

Kevin