[dm-devel] [PATCH v5 3/3] dm mpath: add CONFIG_DM_MULTIPATH_SG_IO - failover for SG_IO

[Paolo Bonzini]

On 05/07/21 15:48, Martin Wilck wrote:
> On Mo, 2021-07-05 at 15:11 +0200, Hannes Reinecke wrote:
>> On 7/5/21 3:02 PM, Paolo Bonzini wrote:
>>> On 02/07/21 16:21, Martin Wilck wrote:
>>>>> SG_IO gives you raw access to the SCSI logic unit, and you get
>>>>> to
>>>>> keep the pieces if anything goes wrong.
>>>>
>>>> That's a very fragile user space API, on the fringe of being
>>>> useless
>>>> IMO.
>>>
>>> Indeed.  If SG_IO is for raw access to an ITL nexus, it shouldn't
>>> be supported at all by mpath devices.  If on the other hand SG_IO is
>>> for raw access to a LUN, there is no reason for it to not support
>>> failover.
>>
>> Or we look at IO_URING_OP_URING_CMD, to send raw cdbs via io_uring.
>> And delegate SG_IO to raw access to an ITL nexus.
>> Doesn't help with existing issues, but should get a clean way
>> forward.
> 
> I still have to understand how this would help with the retrying
> semantics. Wouldn't we get the exact same problem if a path error
> occurs?

Also, how would the URING_CMD API differ from SG_IO modulo one being a 
ioctl and one being io_uring-based?  In the end what you have to do is 
1) send a CDB and optionally some data 2) get back a status and 
optionally some data and sense.  Whether the intended use of the API is 
for an ITL nexus or a LUN doesn't really matter.  So, what is the 
rationale for "SG_IO is for a nexus" in the first place, if you think 
that "raw CDBs for a LUN" is a useful operation?  You can still use 
DM_TABLE_STATUS (iirc) to address a specific ITL nexus if desired.

Besides the virtualization case, think of udev rules that use SG_IO to 
retrieve the device identification page for an mpath device and create 
corresponding symlinks in /dev.  They would fail if the first path is 
not responding, which is not desirable.

Paolo