Re: [Qemu-devel] scsi-hd with discard_granularity and unmap results in Aborted Commands

[Paolo Bonzini <pbonzini@redhat.com>]

Il 19/11/2012 12:49, Stefan Priebe - Profihost AG ha scritto:
>>
>> It still causes hangs no?  Though it works apart from that.
> 
> iscsi/libiscsi and discards works fine since your latest patches:
> 1bd075f29ea6d11853475c7c42734595720c3ac6
> cfb3f5064af2d2e29c976e292c9472dfe9d61e31
> 27cbd828c617944c0f9603763fdf4fa87e7ad923
> b20909195745c34a819aed14ae996b60ab0f591f

Does the console still hang though?

> But in rbd it starts to cancel the discard requests. Both using the same
> guest and host kernel with the same QEMU version.

rbd's cancellation is broken like libiscsi's was.  So the cancellation
succeeds apparently, but it is subject to the same race introduced in
commit 64e69e8 (iscsi: Fix NULL dereferences / races between task
completion and abort, 2012-08-15) for libiscsi.

It risks corruption in case the following happens:

        guest                  qemu                 rbd server
=========================================================================
        send write 1 -------->
                               send write 1 ------>
        cancel write 1 ------>
                               cancel write 1 ---->
           <------------------ cancelled
        send write 2 -------->
                               send write 2 ------>
                                   <--------------- completed write 2
           <------------------ completed write 2
                                   <--------------- completed write 1

Here, the guest would see write 2 superseding write 1, when in fact the
outcome could have been the opposite.  The right behavior is to return
only after the target says whether the cancellation was done or not.
For libiscsi, it was implemented by the commits you mention.

Paolo