Re: [Qemu-devel] [PATCH v7 RFC] block/vxhs: Initial commit to add Veritas HyperScale VxHS block device support

[Stefan Hajnoczi]

On Mon, Nov 28, 2016 at 10:23:41AM +0000, Ketan Nilangekar wrote:
> 
> 
> On 11/25/16, 5:05 PM, "Stefan Hajnoczi" <address@hidden> wrote:
> 
>     On Fri, Nov 25, 2016 at 08:27:26AM +0000, Ketan Nilangekar wrote:
>     > On 11/24/16, 9:38 PM, "Stefan Hajnoczi" <address@hidden> wrote:
>     >     On Thu, Nov 24, 2016 at 11:31:14AM +0000, Ketan Nilangekar wrote:
>     >     > On 11/24/16, 4:41 PM, "Stefan Hajnoczi" <address@hidden> wrote:
>     >     >     On Thu, Nov 24, 2016 at 05:44:37AM +0000, Ketan Nilangekar 
> wrote:
>     >     >     > On 11/24/16, 4:07 AM, "Paolo Bonzini" <address@hidden> 
> wrote:
>     >     >     > >On 23/11/2016 23:09, ashish mittal wrote:
>     >     >     > >> On the topic of protocol security -
>     >     >     > >> 
>     >     >     > >> Would it be enough for the first patch to implement only
>     >     >     > >> authentication and not encryption?
>     >     >     > >
>     >     >     > >Yes, of course.  However, as we introduce more and more 
> QEMU-specific
>     >     >     > >characteristics to a protocol that is already 
> QEMU-specific (it doesn't
>     >     >     > >do failover, etc.), I am still not sure of the actual 
> benefit of using
>     >     >     > >libqnio versus having an NBD server or FUSE driver.
>     >     >     > >
>     >     >     > >You have already mentioned performance, but the design has 
> changed so
>     >     >     > >much that I think one of the two things has to change: 
> either failover
>     >     >     > >moves back to QEMU and there is no (closed source) 
> translator running on
>     >     >     > >the node, or the translator needs to speak a well-known and
>     >     >     > >already-supported protocol.
>     >     >     > 
>     >     >     > IMO design has not changed. Implementation has changed 
> significantly. I would propose that we keep resiliency/failover code out of 
> QEMU driver and implement it entirely in libqnio as planned in a subsequent 
> revision. The VxHS server does not need to understand/handle failover at all. 
>     >     >     > 
>     >     >     > Today libqnio gives us significantly better performance 
> than any NBD/FUSE implementation. We know because we have prototyped with 
> both. Significant improvements to libqnio are also in the pipeline which will 
> use cross memory attach calls to further boost performance. Ofcourse a big 
> reason for the performance is also the HyperScale storage backend but we 
> believe this method of IO tapping/redirecting can be leveraged by other 
> solutions as well.
>     >     >     
>     >     >     By "cross memory attach" do you mean
>     >     >     process_vm_readv(2)/process_vm_writev(2)?
>     >     >   
>     >     > Ketan> Yes.
>     >     >   
>     >     >     That puts us back to square one in terms of security.  You 
> have
>     >     >     (untrusted) QEMU + (untrusted) libqnio directly accessing the 
> memory of
>     >     >     another process on the same machine.  That process is 
> therefore also
>     >     >     untrusted and may only process data for one guest so that 
> guests stay
>     >     >     isolated from each other.
>     >     >     
>     >     > Ketan> Understood but this will be no worse than the current 
> network based communication between qnio and vxhs server. And although we 
> have questions around QEMU trust/vulnerability issues, we are looking to 
> implement basic authentication scheme between libqnio and vxhs server.
>     >     
>     >     This is incorrect.
>     >     
>     >     Cross memory attach is equivalent to ptrace(2) (i.e. debugger) 
> access.
>     >     It means process A reads/writes directly from/to process B memory.  
> Both
>     >     processes must have the same uid/gid.  There is no trust boundary
>     >     between them.
>     >     
>     > Ketan> Not if vxhs server is running as root and initiating the cross 
> mem attach. Which is also why we are proposing a basic authentication 
> mechanism between qemu-vxhs. But anyway the cross memory attach is for a near 
> future implementation. 
>     > 
>     >     Network communication does not require both processes to have the 
> same
>     >     uid/gid.  If you want multiple QEMU processes talking to a single 
> server
>     >     there must be a trust boundary between client and server.  The 
> server
>     >     can validate the input from the client and reject undesired 
> operations.
>     > 
>     > Ketan> This is what we are trying to propose. With the addition of 
> authentication between qemu-vxhs server, we should be able to achieve this. 
> Question is, would that be acceptable?
>     >     
>     >     Hope this makes sense now.
>     >     
>     >     Two architectures that implement the QEMU trust model correctly are:
>     >     
>     >     1. Cross memory attach: each QEMU process has a dedicated vxhs 
> server
>     >        process to prevent guests from attacking each other.  This is 
> where I
>     >        said you might as well put the code inside QEMU since there is no
>     >        isolation anyway.  From what you've said it sounds like the vxhs
>     >        server needs a host-wide view and is responsible for all guests
>     >        running on the host, so I guess we have to rule out this
>     >        architecture.
>     >     
>     >     2. Network communication: one vxhs server process and multiple 
> guests.
>     >        Here you might as well use NBD or iSCSI because it already 
> exists and
>     >        the vxhs driver doesn't add any unique functionality over 
> existing
>     >        protocols.
>     > 
>     > Ketan> NBD does not give us the performance we are trying to achieve. 
> Besides NBD does not have any authentication support.
>     
>     NBD over TCP supports TLS with X.509 certificate authentication.  I
>     think Daniel Berrange mentioned that.
> 
> Ketan> I saw the patch to nbd that was merged in 2015. Before that NBD did 
> not have any auth as Daniel Berrange mentioned. 
>     
>     NBD over AF_UNIX does not need authentication because it relies on file
>     permissions for access control.  Each guest should have its own UNIX
>     domain socket that it connects to.  That socket can only see exports
>     that have been assigned to the guest.
>     
>     > There is a hybrid 2.a approach which uses both 1 & 2 but I’d keep that 
> for a later discussion.
>     
>     Please discuss it now so everyone gets on the same page.  I think there
>     is a big gap and we need to communicate so that progress can be made.
> 
> Ketan> The approach was to use cross mem attach for IO path and a simplified 
> network IO lib for resiliency/failover. Did not want to derail the current 
> discussion hence the suggestion to take it up later.

Why does the client have to know about failover if it's connected to a
server process on the same host?  I thought the server process manages
networking issues (like the actual protocol to speak to other VxHS nodes
and for failover).

>     >     >     There's an easier way to get even better performance: get rid 
> of libqnio
>     >     >     and the external process.  Move the code from the external 
> process into
>     >     >     QEMU to eliminate the 
> process_vm_readv(2)/process_vm_writev(2) and
>     >     >     context switching.
>     >     >     
>     >     >     Can you remind me why there needs to be an external process?
>     >     >  
>     >     > Ketan>  Apart from virtualizing the available direct attached 
> storage on the compute, vxhs storage backend (the external process) provides 
> features such as storage QoS, resiliency, efficient use of direct attached 
> storage, automatic storage recovery points (snapshots) etc. Implementing this 
> in QEMU is not practical and not the purpose of proposing this driver.
>     >     
>     >     This sounds similar to what QEMU and Linux (file systems, LVM, RAID,
>     >     etc) already do.  It brings to mind a third architecture:
>     >     
>     >     3. A Linux driver or file system.  Then QEMU opens a raw block 
> device.
>     >        This is what the Ceph rbd block driver in Linux does.  This
>     >        architecture has a kernel-userspace boundary so vxhs does not 
> have to
>     >        trust QEMU.
>     >     
>     >     I suggest Architecture #2.  You'll be able to deploy on existing 
> systems
>     >     because QEMU already supports NBD or iSCSI.  Use the time you gain 
> from
>     >     switching to this architecture on benchmarking and optimizing NBD or
>     >     iSCSI so performance is closer to your goal.
>     >     
>     > Ketan> We have made a choice to go with QEMU driver approach after 
> serious evaluation of most if not all standard IO tapping mechanisms 
> including NFS, NBD and FUSE. None of these has been able to deliver the 
> performance that we have set ourselves to achieve. Hence the effort to 
> propose this new IO tap which we believe will provide an alternate to the 
> existing mechanisms and hopefully benefit the community.
>     
>     I thought the VxHS block driver was another network block driver like
>     GlusterFS or Sheepdog but you are actually proposing a new local I/O tap
>     with the goal of better performance.
> 
> Ketan> The VxHS block driver is a new local IO tap with the goal of better 
> performance specifically when used with the VxHS server. This coupled with 
> shared mem IPC (like cross mem attach) could be a much better IO tap option 
> for qemu users. This will also avoid context switch between qemu/network 
> stack to service which happens today in NBD.
> 
>     
>     Please share fio(1) or other standard benchmark configuration files and
>     performance results.
> 
> Ketan> We have fio results with the VxHS storage backend which I am not sure 
> I can share in a public forum. 
>     
>     NBD and libqnio wire protocols have comparable performance
>     characteristics.  There is no magic that should give either one a
>     fundamental edge over the other.  Am I missing something?
> 
> Ketan> I have not seen the NBD code but few things which we considered and 
> are part of libqnio (though not exclusively) are low protocol overhead, 
> threading model, queueing, latencies, memory pools, zero data copies in 
> user-land, scatter-gather write/read etc. Again these are not exclusive to 
> libqnio but could give one protocol the edge over the other. Also part of the 
> “magic” is also in the VxHS storage backend which is able to ingest the IOs 
> with lower latencies.
>     
>     The main performance difference is probably that libqnio opens 8
>     simultaneous connections but that's not unique to the wire protocol.
>     What happens when you run 8 NBD simultaneous TCP connections?
> 
> Ketan> Possibly. We have not benchmarked this.

There must be benchmark data if you want to add a new feature or modify
existing code for performance reasons.  This rule is followed in QEMU so
that performance changes are justified.

I'm afraid that when you look into the performance you'll find that any
performance difference between NBD and this VxHS patch series is due to
implementation differences that can be ported across to QEMU NBD, rather
than wire protocol differences.

If that's the case then it would save a lot of time to use NBD over
AF_UNIX for now.  You could focus efforts on achieving the final
architecture you've explained with cross memory attach.

Please take a look at vhost-user-scsi, which folks from Nutanix are
currently working on.  See "[PATCH v2 0/3] Introduce vhost-user-scsi and
sample application" on qemu-devel.  It is a true zero-copy local I/O tap
because it shares guest RAM.  This is more efficient than cross memory
attach's single memory copy.  It does not require running the server as
root.  This is the #1 thing you should evaluate for your final
architecture.

vhost-user-scsi works on the virtio-scsi emulation level.  That means
the server must implement the virtio-scsi vring and device emulation.
It is not a block driver.  By hooking in at this level you can achieve
the best performance but you lose all QEMU block layer functionality and
need to implement your own SCSI target.  You also need to consider live
migration.

Stefan

signature.asc
Description: PGP signature