Re: [PATCH v6 8/8] eni_vdpa: add vDPA driver for Alibaba ENI

From: Wu Zongyong
Date: Wed Oct 27 2021 - 03:04:30 EST

Next message: Zqiang: "Re: [PATCH] io-wq: Remove unnecessary rcu_read_lock/unlock() in raw spinlock critical section"
Previous message: Peter Xu: "Re: [PATCH v1] mm, pagemap: expose hwpoison entry"
In reply to: Jason Wang: "Re: [PATCH v6 8/8] eni_vdpa: add vDPA driver for Alibaba ENI"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]

On Wed, Oct 27, 2021 at 11:55:05AM +0800, Jason Wang wrote:
> On Wed, Oct 27, 2021 at 10:47 AM Wu Zongyong
> <wuzongyong@xxxxxxxxxxxxxxxxx> wrote:
> >
> > On Mon, Oct 25, 2021 at 12:40:41PM +0800, Jason Wang wrote:
> > >
> > > 在 2021/10/25 上午11:21, Wu Zongyong 写道:
> > > > On Mon, Oct 25, 2021 at 10:27:31AM +0800, Jason Wang wrote:
> > > > > On Fri, Oct 22, 2021 at 10:44 AM Wu Zongyong
> > > > > <wuzongyong@xxxxxxxxxxxxxxxxx> wrote:
> > > > > > This patch adds a new vDPA driver for Alibaba ENI(Elastic Network
> > > > > > Interface) which is build upon virtio 0.9.5 specification.
> > > > > > And this driver doesn't support to run on BE host.
> > > > > >
> > > > > > Signed-off-by: Wu Zongyong <wuzongyong@xxxxxxxxxxxxxxxxx>
> > > > > > ---
> > > > > > drivers/vdpa/Kconfig | 8 +
> > > > > > drivers/vdpa/Makefile | 1 +
> > > > > > drivers/vdpa/alibaba/Makefile | 3 +
> > > > > > drivers/vdpa/alibaba/eni_vdpa.c | 553 ++++++++++++++++++++++++++++++++
> > > > > > 4 files changed, 565 insertions(+)
> > > > > > create mode 100644 drivers/vdpa/alibaba/Makefile
> > > > > > create mode 100644 drivers/vdpa/alibaba/eni_vdpa.c
> > > > > >
> > > > > > diff --git a/drivers/vdpa/Kconfig b/drivers/vdpa/Kconfig
> > > > > > index 3d91982d8371..c0232a2148a7 100644
> > > > > > --- a/drivers/vdpa/Kconfig
> > > > > > +++ b/drivers/vdpa/Kconfig
> > > > > > @@ -78,4 +78,12 @@ config VP_VDPA
> > > > > > help
> > > > > > This kernel module bridges virtio PCI device to vDPA bus.
> > > > > >
> > > > > > +config ALIBABA_ENI_VDPA
> > > > > > + tristate "vDPA driver for Alibaba ENI"
> > > > > > + select VIRTIO_PCI_LEGACY_LIB
> > > > > > + depends on PCI_MSI && !CPU_BIG_ENDIAN
> > > > > > + help
> > > > > > + VDPA driver for Alibaba ENI(Elastic Network Interface) which is build upon
> > > > > > + virtio 0.9.5 specification.
> > > > > > +
> > > > > > endif # VDPA
> > > > > > diff --git a/drivers/vdpa/Makefile b/drivers/vdpa/Makefile
> > > > > > index f02ebed33f19..15665563a7f4 100644
> > > > > > --- a/drivers/vdpa/Makefile
> > > > > > +++ b/drivers/vdpa/Makefile
> > > > > > @@ -5,3 +5,4 @@ obj-$(CONFIG_VDPA_USER) += vdpa_user/
> > > > > > obj-$(CONFIG_IFCVF) += ifcvf/
> > > > > > obj-$(CONFIG_MLX5_VDPA) += mlx5/
> > > > > > obj-$(CONFIG_VP_VDPA) += virtio_pci/
> > > > > > +obj-$(CONFIG_ALIBABA_ENI_VDPA) += alibaba/
> > > > > > diff --git a/drivers/vdpa/alibaba/Makefile b/drivers/vdpa/alibaba/Makefile
> > > > > > new file mode 100644
> > > > > > index 000000000000..ef4aae69f87a
> > > > > > --- /dev/null
> > > > > > +++ b/drivers/vdpa/alibaba/Makefile
> > > > > > @@ -0,0 +1,3 @@
> > > > > > +# SPDX-License-Identifier: GPL-2.0
> > > > > > +obj-$(CONFIG_ALIBABA_ENI_VDPA) += eni_vdpa.o
> > > > > > +
> > > > > > diff --git a/drivers/vdpa/alibaba/eni_vdpa.c b/drivers/vdpa/alibaba/eni_vdpa.c
> > > > > > new file mode 100644
> > > > > > index 000000000000..6a09f157d810
> > > > > > --- /dev/null
> > > > > > +++ b/drivers/vdpa/alibaba/eni_vdpa.c
> > > > > > @@ -0,0 +1,553 @@
> > > > > > +// SPDX-License-Identifier: GPL-2.0-only
> > > > > > +/*
> > > > > > + * vDPA bridge driver for Alibaba ENI(Elastic Network Interface)
> > > > > > + *
> > > > > > + * Copyright (c) 2021, Alibaba Inc. All rights reserved.
> > > > > > + * Author: Wu Zongyong <wuzongyong@xxxxxxxxxxxxxxxxx>
> > > > > > + *
> > > > > > + */
> > > > > > +
> > > > > > +#include "linux/bits.h"
> > > > > > +#include <linux/interrupt.h>
> > > > > > +#include <linux/module.h>
> > > > > > +#include <linux/pci.h>
> > > > > > +#include <linux/vdpa.h>
> > > > > > +#include <linux/virtio.h>
> > > > > > +#include <linux/virtio_config.h>
> > > > > > +#include <linux/virtio_ring.h>
> > > > > > +#include <linux/virtio_pci.h>
> > > > > > +#include <linux/virtio_pci_legacy.h>
> > > > > > +#include <uapi/linux/virtio_net.h>
> > > > > > +
> > > > > > +#define ENI_MSIX_NAME_SIZE 256
> > > > > > +
> > > > > > +#define ENI_ERR(pdev, fmt, ...) \
> > > > > > + dev_err(&pdev->dev, "%s"fmt, "eni_vdpa: ", ##__VA_ARGS__)
> > > > > > +#define ENI_DBG(pdev, fmt, ...) \
> > > > > > + dev_dbg(&pdev->dev, "%s"fmt, "eni_vdpa: ", ##__VA_ARGS__)
> > > > > > +#define ENI_INFO(pdev, fmt, ...) \
> > > > > > + dev_info(&pdev->dev, "%s"fmt, "eni_vdpa: ", ##__VA_ARGS__)
> > > > > > +
> > > > > > +struct eni_vring {
> > > > > > + void __iomem *notify;
> > > > > > + char msix_name[ENI_MSIX_NAME_SIZE];
> > > > > > + struct vdpa_callback cb;
> > > > > > + int irq;
> > > > > > +};
> > > > > > +
> > > > > > +struct eni_vdpa {
> > > > > > + struct vdpa_device vdpa;
> > > > > > + struct virtio_pci_legacy_device ldev;
> > > > > > + struct eni_vring *vring;
> > > > > > + struct vdpa_callback config_cb;
> > > > > > + char msix_name[ENI_MSIX_NAME_SIZE];
> > > > > > + int config_irq;
> > > > > > + int queues;
> > > > > > + int vectors;
> > > > > > +};
> > > > > > +
> > > > > > +static struct eni_vdpa *vdpa_to_eni(struct vdpa_device *vdpa)
> > > > > > +{
> > > > > > + return container_of(vdpa, struct eni_vdpa, vdpa);
> > > > > > +}
> > > > > > +
> > > > > > +static struct virtio_pci_legacy_device *vdpa_to_ldev(struct vdpa_device *vdpa)
> > > > > > +{
> > > > > > + struct eni_vdpa *eni_vdpa = vdpa_to_eni(vdpa);
> > > > > > +
> > > > > > + return &eni_vdpa->ldev;
> > > > > > +}
> > > > > > +
> > > > > > +static u64 eni_vdpa_get_features(struct vdpa_device *vdpa)
> > > > > > +{
> > > > > > + struct virtio_pci_legacy_device *ldev = vdpa_to_ldev(vdpa);
> > > > > > + u64 features = vp_legacy_get_features(ldev);
> > > > > > +
> > > > > > + features |= BIT_ULL(VIRTIO_F_ACCESS_PLATFORM);
> > > > > > + features |= BIT_ULL(VIRTIO_F_ORDER_PLATFORM);
> > > > > > +
> > > > > > + return features;
> > > > > > +}
> > > > > > +
> > > > > > +static int eni_vdpa_set_features(struct vdpa_device *vdpa, u64 features)
> > > > > > +{
> > > > > > + struct virtio_pci_legacy_device *ldev = vdpa_to_ldev(vdpa);
> > > > > > +
> > > > > > + if (!(features & BIT_ULL(VIRTIO_NET_F_MRG_RXBUF)) && features) {
> > > > > > + ENI_ERR(ldev->pci_dev,
> > > > > > + "VIRTIO_NET_F_MRG_RXBUF is not negotiated\n");
> > > > > > + return -EINVAL;
> > > > > > + }
> > > > > > +
> > > > > > + vp_legacy_set_features(ldev, (u32)features);
> > > > > > +
> > > > > > + return 0;
> > > > > > +}
> > > > > > +
> > > > > So my comments have not been addressed since v4. Please address or
> > > > > answer the questions before posting a new version.
> > > > >
> > > > > Thanks
> > > > Sorry, I forgot to reply the comments on this patch.
> > > >
> > > >
> > > > > > +static u64 eni_vdpa_get_features(struct vdpa_device *vdpa)
> > > > > > +{
> > > > > > + struct virtio_pci_legacy_device *ldev = vdpa_to_ldev(vdpa);
> > > > > > + u64 features = vp_legacy_get_features(ldev);
> > > > > > +
> > > > > > + features |= BIT_ULL(VIRTIO_F_ACCESS_PLATFORM);
> > > > > > + features |= BIT_ULL(VIRTIO_F_ORDER_PLATFORM);
> > > > > VERSION_1 is also needed?
> > > > >
> > > > No, queue align of legacy devices should be 4096,
> > >
> > >
> > > Let's use VIRTIO_PCI_VRING_ALIGN instead of PAGE_SIZE in get_vq_align then
> > > since PAGE_SIZE is not necessarily 4096.
> >
> > I will fix it.
> > >
> > >
> > > > but queue align of
> > > > devices with VERSION_1 are SMP_CACHE_BYTES which may not equals to
> > > > 4096.
> > > > If we set the VERSION_1, ENI will not work due to the queue align.
> > >
> > >
> > > Interesting, so I think it can only be used with legacy virtio drivers in
> > > the guest.
> > Yes, ENI can only work well with legacy virtio drivers in the guest from
> > my test.
> >
> > >
> > > One major drawbacks is that guest can only see 32 feature bits which means
> > > we can't advertise VIRTIO_F_ACCESS_PLATFORM and VIRTIO_F_ORDER_PLATFORM to
> > > guest:
> > >
> > > /* virtio config->get_features() implementation */
> > > static u64 vp_get_features(struct virtio_device *vdev)
> > > {
> > > struct virtio_pci_device *vp_dev = to_vp_device(vdev);
> > >
> > > /* When someone needs more than 32 feature bits, we'll need to
> > > * steal a bit to indicate that the rest are somewhere else. */
> > > return ioread32(vp_dev->ioaddr + VIRTIO_PCI_HOST_FEATURES);
> > > }
> > >
> > > For VIRTIO_F_ACCESS_PLATFORM, it should be fine. But how about
> > > VIRTIO_F_ORDER_PLATFORM?
> > >
> > Is it necessary to advertise these two bits?
> > ENI works well without these two bits regardless of bare metal or virtual
> > machine.
>
> For VIRTIO_F_ACCESS_PLATFORM. When using with virtio-vDPA it should be
> fine since it will be negotiated. When using with vhost-vDPA, it can't
> be advertised to the guest, so vIOMMU will be bypassed.
>
> For VITIO_F_ORDER_PLATFORM, it's basically a question that whether
> platform specific barrier (DMA barrier) is needed, see virtio_rmb():
>
> static inline void virtio_rmb(bool weak_barriers)
> {
> if (weak_barriers)
> virt_rmb();
> else
> dma_rmb();
> }
>
> So it looks to me x86 should be fine, but for others platform it might
> not. Maybe we can enable this driver for x86 first.

It's okay for me to enable it only for x86 via Kconfig since ENI can
only be used in x86 machines on Alibaba Cloud for now.

>
> Thanks
>
> > >
> > > >
> > > > > > +
> > > > > > + return features;
> > > > > > +}
> > > > > > +
> > > > > > +static int eni_vdpa_set_features(struct vdpa_device *vdpa, u64 features)
> > > > > > +{
> > > > > > + struct virtio_pci_legacy_device *ldev = vdpa_to_ldev(vdpa);
> > > > > > +
> > > > > > + if (!(features & BIT_ULL(VIRTIO_NET_F_MRG_RXBUF)) && features) {
> > > > > > + ENI_ERR(ldev->pci_dev,
> > > > > > + "VIRTIO_NET_F_MRG_RXBUF is not negotiated\n");
> > > > > > + return -EINVAL;
> > > > > Do we need to make sure FEATURE_OK is not set in this case or the ENI can do
> > > > > this for us?
> > > > Why we need to check this? I don't get what you worried about.
> > >
> > >
> > > I thought the plan is to advertise the VERSION_1, so failing when without
> > > mrg_rxbuf is a must. But looks like I was wrong, and there's no need to
> > > mandate mrg_rxbuf in future versions.
> >
> > I think it is necessary to check F_MRG_RXBUF here, since ENI will not
> > work when use qemu with mrx_rxbuf=off.
> > >
> > > Thanks
> > >
> > >
> > > >
> > > > > Other looks good.
> > > > >
> > > > > Thanks
> >

Next message: Zqiang: "Re: [PATCH] io-wq: Remove unnecessary rcu_read_lock/unlock() in raw spinlock critical section"
Previous message: Peter Xu: "Re: [PATCH v1] mm, pagemap: expose hwpoison entry"
In reply to: Jason Wang: "Re: [PATCH v6 8/8] eni_vdpa: add vDPA driver for Alibaba ENI"
Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]