On Fri, Dec 18, 2015 at 10:01:25AM -0800, Rich Lane wrote:
> I'm using the vhost callbacks and struct virtio_net with the vhost PMD in a
> few
> ways:
Rich, thanks for the info!
>
> 1. new_device/destroy_device: Link state change (will be covered by the link
> status interrupt).
> 2. new_device: Add first queue to datapath.
I'm wondering why vring_state_changed() is not used, as it will also be
triggered at the beginning, when the default queue (the first queue) is
enabled.
> 3. vring_state_changed: Add/remove queue to datapath.
> 4. destroy_device: Remove all queues (vring_state_changed is not called when
> qemu is killed).
I had a plan to invoke vring_state_changed() to disable all vrings
when destroy_device() is called.
> 5. new_device and struct virtio_net: Determine NUMA node of the VM.
You can get the 'struct virtio_net' dev from all above callbacks.
>
> The vring_state_changed callback is necessary because the VM might not be
> using
> the maximum number of RX queues. If I boot Linux in the VM it will start out
> using one RX queue, which can be changed with ethtool. The DPDK app in the
> host
> needs to be notified that it can start sending traffic to the new queue.
>
> The vring_state_changed callback is also useful for guest TX queues to avoid
> reading from an inactive queue.
>
> API I'd like to have:
>
> 1. Link status interrupt.
To vhost pmd, new_device()/destroy_device() equals to the link status
interrupt, where new_device() is a link up, and destroy_device() is link
down().
> 2. New queue_state_changed callback. Unlike vring_state_changed this should
> cover the first queue at new_device and removal of all queues at
> destroy_device.
As stated above, vring_state_changed() should be able to do that, except
the one on destroy_device(), which is not done yet.
> 3. Per-queue or per-device NUMA node info.
You can query the NUMA node info implicitly by get_mempolicy(); check
numa_realloc() at lib/librte_vhost/virtio-net.c for reference.
--yliu
>
> On Thu, Dec 17, 2015 at 8:28 PM, Tetsuya Mukawa <mukawa at igel.co.jp> wrote:
>
> On 2015/12/18 13:15, Yuanhan Liu wrote:
> > On Fri, Dec 18, 2015 at 12:15:42PM +0900, Tetsuya Mukawa wrote:
> >> On 2015/12/17 20:42, Yuanhan Liu wrote:
> >>> On Tue, Nov 24, 2015 at 06:00:01PM +0900, Tetsuya Mukawa wrote:
> >>>> The vhost PMD will be a wrapper of vhost library, but some of vhost
> >>>> library APIs cannot be mapped to ethdev library APIs.
> >>>> Becasue of this, in some cases, we still need to use vhost library
> APIs
> >>>> for a port created by the vhost PMD.
> >>>>
> >>>> Currently, when virtio device is created and destroyed, vhost library
> >>>> will call one of callback handlers. The vhost PMD need to use this
> >>>> pair of callback handlers to know which virtio devices are connected
> >>>> actually.
> >>>> Because we can register only one pair of callbacks to vhost library,
> if
> >>>> the PMD use it, DPDK applications cannot have a way to know the
> events.
> >>>>
> >>>> This may break legacy DPDK applications that uses vhost library. To
> prevent
> >>>> it, this patch adds one more pair of callbacks to vhost library
> especially
> >>>> for the vhost PMD.
> >>>> With the patch, legacy applications can use the vhost PMD even if
> they
> need
> >>>> additional specific handling for virtio device creation and
> destruction.
> >>>>
> >>>> For example, legacy application can call
> >>>> rte_vhost_enable_guest_notification() in callbacks to change setting.
> >>> TBH, I never liked it since the beginning. Introducing two callbacks
> >>> for one event is a bit messy, and therefore error prone.
> >> I agree with you.
> >>
> >>> I have been thinking this occasionally last few weeks, and have came
> >>> up something that we may introduce another layer callback based on
> >>> the vhost pmd itself, by a new API:
> >>>
> >>>? ? ?rte_eth_vhost_register_callback().
> >>>
> >>> And we then call those new callback inside the vhost pmd new_device()
> >>> and vhost pmd destroy_device() implementations.
> >>>
> >>> And we could have same callbacks like vhost have, but I'm thinking
> >>> that new_device() and destroy_device() doesn't sound like a good name
> >>> to a PMD driver. Maybe a name like "link_state_changed" is better?
> >>>
> >>> What do you think of that?
> >> Yes,? "link_state_changed" will be good.
> >>
> >> BTW, I thought it was ok that an DPDK app that used vhost PMD called
> >> vhost library APIs directly.
> >> But probably you may feel strangeness about it. Is this correct?
> > Unluckily, that's true :)
> >
> >> If so, how about implementing legacy status interrupt mechanism to
> vhost
> >> PMD?
> >> For example, an DPDK app can register callback handler like
> >> "examples/link_status_interrupt".
> >>
> >> Also, if the app doesn't call vhost library APIs directly,
> >> rte_eth_vhost_portid2vdev() will be needless, because the app doesn't
> >> need to handle virtio device structure anymore.
> >>
> >>>
> >>> On the other hand, I'm still thinking is that really necessary to let
> >>> the application be able to call vhost functions like
> rte_vhost_enable_guest_notification()
> >>> with the vhost PMD driver?
> >> Basic concept of my patch is that vhost PMD will provides the features
> >> that vhost library provides.
> > I don't think that's necessary. Let's just treat it as a normal pmd
> > driver, having nothing to do with vhost library.
> >
> >> How about removing rte_vhost_enable_guest_notification() from "vhost
> >> library"?
> >> (I also not sure what are use cases)
> >> If we can do this, vhost PMD also doesn't need to take care of it.
> >> Or if rte_vhost_enable_guest_notification() will be removed in the
> >> future, vhost PMD is able to ignore it.
> > You could either call it in vhost-pmd (which you already have done
> that),
> > or ignore it in vhost-pmd, but dont' remove it from vhost library.
> >
> >> Please let me correct up my thinking about your questions.
> >>? - Change concept of patch not to call vhost library APIs directly.
> >> These should be wrapped by ethdev APIs.
> >>? - Remove rte_eth_vhost_portid2vdev(), because of above concept
> changing.
> >>? - Implement legacy status changed interrupt to vhost PMD instead of
> >> using own callback mechanism.
> >>? - Check if we can remove rte_vhost_enable_guest_notification() from
> >> vhost library.
> > So, how about making it __fare__ simple as the first step, to get merged
> > easily, that we don't assume the applications will call any vhost
> library
> > functions any more, so that we don't need the callback, and we don't
> need
> > the rte_eth_vhost_portid2vdev(), either. Again, just let it be a fare
> > normal (nothing special) pmd driver.? (UNLESS, there is a real must,
> which
> > I don't see so far).
> >
> > Tetsuya, what do you think of that then?
>
> I agree with you. But will wait a few days.
> Because if someone wants to use it from vhost PMD, they probably will
> provides use cases.
> And if there are no use cases, let's do like above.
>
> Thanks,
> Tetsuya
>
>
