Re: [Qemu-devel] [PATCH 0/1] vhost-user: Add a protocol extension for client responses to vhost commands.

[Michael S. Tsirkin]

On Sat, Jun 25, 2016 at 03:13:54AM +0000, Prerna Saxena wrote:
> 
> 
> 
> 
> 
> On 25/06/16 4:43 am, "Michael S. Tsirkin" <address@hidden> wrote:
> 
> >On Fri, Jun 24, 2016 at 05:39:31PM +0000, Prerna Saxena wrote:
> >> 
> >> 
> >> On 24/06/16 9:15 pm, "Felipe Franciosi" <address@hidden> wrote:
> >> 
> >> >We talked to MST on IRC a while back and he brainstormed the idea of 
> >> >doing this per-message.
> >> >(I even recall proposing to call this feature REPLY_ALL and he suggested 
> >> >REPLY_ANY due to that.)
> >> >
> >> >I agree with doing it per message, as the protocol itself should be 
> >> >flexible in that sense.
> >> >(Even if qemu today will probably want to ask for a reply in all 
> >> >messages.)
> >> 
> >> In fact, the current implementation does exactly this. If 
> >> VHOST_USER_PROTOCOL_F_REPLY_ACK is negotiated, the current QEMU patch sets 
> >> the NEED_RESPONSE flag bit for all outgoing messages — basically enforcing 
> >> the vhost-user application to respond to all messages.
> >
> >
> >This seems unnecessary. Let's only do that for messages that actually
> >need to be synchronous.
> 
> It would be nice to distinguish the vhost-user protocol itself from its QEMU 
> implementation.
> The protocol should, in theory, have provision for an implementation (such as 
> QEMU’s vhost-user implementation) to seek response for _any_ command. 
> However, we can choose to be selective in our QEMU implementation and just 
> have limited commands currently send a response, such as SET_MEM_TABLE. 
> 
> In other words, we will still require the NEED_RESPONSE flag bit defined, but 
> we can just set it to 1 it for SET_MEM_TABLE command in our QEMU 
> implementation. All other vhost-user commands are sent from QEMU setting this 
> to 0, so the application does not send an ack.
> 
> Michael, Does that correctly summarize what you were meaning to suggest here ?
> 
> Regards,
> Prerna

Exactly.

> 
> >
> >> >
> >> >On 24/06/2016, 14:59, "Qemu-devel on behalf of Marc-André Lureau" 
> >> ><address@hidden on behalf of address@hidden> wrote:
> >> >
> >> >Hi
> >> >
> >> >On Fri, Jun 24, 2016 at 10:17 AM, Prerna Saxena <address@hidden> wrote:
> >> >> From: Prerna Saxena <address@hidden>
> >> >>
> >> >> The current vhost-user protocol requires the client to send responses 
> >> >> to only few commands. For the remaining commands, it is impossible for 
> >> >> QEMU to know the status of the requested operation -- ie, did it 
> >> >> succeed at all, and if so, at what time.
> >> >>
> >> >> This is inconvenient, and can also lead to races. As an example:
> >> >>
> >> >> (1) qemu sends a SET_MEM_TABLE to the backend (eg, a vhost-user net 
> >> >> application) and SET_MEM_TABLE doesn't require a reply according to the 
> >> >> spec.
> >> >> (2) qemu commits the memory to the guest.
> >> >> (3) guest issues an I/O operation over a new memory region which was 
> >> >> configured on (1)
> >> >> (4) The application hasn't yet remapped the memory, but it sees the I/O 
> >> >> request.
> >> >> (5) The application cannot satisfy the request because it doesn't know 
> >> >> about those GPAs
> >> >>
> >> >> Note that the kernel implementation does not suffer from this 
> >> >> limitation since messages are sent via an ioctl(). The ioctl() blocks 
> >> >> until the backend (eg. vhost-net) completes the command and returns 
> >> >> (with an error code).
> >> >>
> >> >> Changing the behaviour of current vhost-user commands would break 
> >> >> existing applications. This patch introduces a protocol extension, 
> >> >> VHOST_USER_PROTOCOL_F_REPLY_ACK. This feature, if negotiated, allows 
> >> >> QEMU to annotate messages to the application that it seeks a response 
> >> >> for. The application must then respond to qemu by providing a status 
> >> >> about the requested operation.
> >> >
> >> >I like the idea, as I encountered a similar issue in my
> >> >"vhost-user-gpu" development (which I worked around by sending a dump
> >> >GET_FEATURES.. to sync things). But I question the need to have a flag
> >> >per message. I think if the protocol feature is negociated, all
> >> >messages should have a reply. Why do you want it to be per-message?
> >> >
> >> >thanks
> >> >
> >> >-- 
> >> >Marc-André Lureau
> >> >
> >> >
> >> >