Re: [PATCH 00/10] vhost/qemu: thread per IO SCSI vq

[Stefan Hajnoczi <stefanha@gmail.com>]

On Thu, Nov 19, 2020 at 4:13 PM Mike Christie
<michael.christie@oracle.com> wrote:
>
> On 11/19/20 8:46 AM, Michael S. Tsirkin wrote:
> > On Wed, Nov 18, 2020 at 11:31:17AM +0000, Stefan Hajnoczi wrote:
> >>> My preference has been:
> >>>
> >>> 1. If we were to ditch cgroups, then add a new interface that would allow
> >>> us to bind threads to a specific CPU, so that it lines up with the guest's
> >>> mq to CPU mapping.
> >>
> >> A 1:1 vCPU/vq->CPU mapping isn't desirable in all cases.
> >>
> >> The CPU affinity is a userspace policy decision. The host kernel should
> >> provide a mechanism but not the policy. That way userspace can decide
> >> which workers are shared by multiple vqs and on which physical CPUs they
> >> should run.
> >
> > So if we let userspace dictate the threading policy then I think binding
> > vqs to userspace threads and running there makes the most sense,
> > no need to create the threads.
> >
>
> Just to make sure I am on the same page, in one of the first postings of
> this set at the bottom of the mail:
>
> https://www.spinics.net/lists/linux-scsi/msg148322.html
>
> I asked about a new interface and had done something more like what
> Stefan posted:
>
>    struct vhost_vq_worker_info {
>        /*
>         * The pid of an existing vhost worker that this vq will be
>         * assigned to. When pid is 0 the virtqueue is assigned to the
>         * default vhost worker. When pid is -1 a new worker thread is
>         * created for this virtqueue. When pid is -2 the virtqueue's
>         * worker thread is unchanged.
>         *
>         * If a vhost worker no longer has any virtqueues assigned to it
>         * then it will terminate.
>         *
>         * The pid of the vhost worker is stored to this field when the
>         * ioctl completes successfully. Use pid -2 to query the current
>         * vhost worker pid.
>         */
>        __kernel_pid_t pid;  /* in/out */
>
>        /* The virtqueue index*/
>        unsigned int vq_idx; /* in */
>    };
>
> This approach is simple and it allowed me to have userspace map queues
> and threads optimally for our setups.
>
> Note: Stefan, in response to your previous comment, I am just using my
> 1:1 mapping as an example and would make it configurable from userspace.
>
> In the email above are you guys suggesting to execute the SCSI/vhost
> requests in userspace? We should not do that because:
>
> 1. It negates part of what makes vhost fast where we do not have to kick
> out to userspace then back to the kernel.
>
> 2. It's not doable or becomes a crazy mess because vhost-scsi is tied to
> the scsi/target layer in the kernel. You can't process the scsi command
> in userspace since the scsi state machine and all its configuration info
> is in the kernel's scsi/target layer.
>
> For example, I was just the maintainer of the target_core_user module
> that hooks into LIO/target on the backend (vhost-scsi hooks in on the
> front end) and passes commands to userspace and there we have a
> semi-shadow state machine. It gets nasty to try and maintain/sync state
> between lio/target core in the kernel and in userspace. We also see the
> perf loss I mentioned in #1.

No, if I understand Michael correctly he has suggested a different approach.

My suggestion was that the kernel continues to manage the worker
threads but an ioctl allows userspace to control the policy.

I think Michael is saying that the kernel shouldn't manage/create
threads. Userspace should create threads and then invoke an ioctl from
those threads.

The ioctl will call into the vhost driver where it will execute
something similar to vhost_worker(). So this ioctl will block while
the kernel is using the thread to process vqs.

What isn't clear to me is how to tell the kernel which vqs are
processed by a thread. We could try to pass that information into the
ioctl. I'm not sure what the cleanest solution is here.

Maybe something like:

struct vhost_run_worker_info {
    struct timespec *timeout;
    sigset_t *sigmask;

    /* List of virtqueues to process */
    unsigned nvqs;
    unsigned vqs[];
};

/* This blocks until the timeout is reached, a signal is received, or
the vhost device is destroyed */
int ret = ioctl(vhost_fd, VHOST_RUN_WORKER, &info);

As you can see, userspace isn't involved with dealing with the
requests. It just acts as a thread donor to the vhost driver.

We would want the VHOST_RUN_WORKER calls to be infrequent to avoid the
penalty of switching into the kernel, copying in the arguments, etc.

Michael: is this the kind of thing you were thinking of?

Stefan