On Fri, Nov 11, 2022 at 03:17:14PM +0800, Jason Wang
wrote:> On Thu, Nov 10, 2022 at 7:01 PM Michael S. Tsirkin <mst at
redhat.com> wrote:
> >
> > On Thu, Nov 10, 2022 at 03:58:21PM +0800, Jason Wang wrote:
> > > This patch allows device features to be provisioned via vdpa.
This
> > > will be useful for preserving migration compatibility between
source
> > > and destination:
> > >
> > > # vdpa dev add name dev1 mgmtdev pci/0000:02:00.0 device_features
0x300020000
> > > # dev1: mac 52:54:00:12:34:56 link up link_announce false mtu
65535
> > > negotiated_features CTRL_VQ VERSION_1 ACCESS_PLATFORM
> > >
> > > Signed-off-by: Jason Wang <jasowang at redhat.com>
> > > ---
> > > man/man8/vdpa-dev.8 | 10 ++++++++++
> > > vdpa/include/uapi/linux/vdpa.h | 1 +
> > > vdpa/vdpa.c | 27 ++++++++++++++++++++++++++-
> > > 3 files changed, 37 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/man/man8/vdpa-dev.8 b/man/man8/vdpa-dev.8
> > > index 9faf3838..bb45b4a6 100644
> > > --- a/man/man8/vdpa-dev.8
> > > +++ b/man/man8/vdpa-dev.8
> > > @@ -31,6 +31,7 @@ vdpa-dev \- vdpa device configuration
> > > .I NAME
> > > .B mgmtdev
> > > .I MGMTDEV
> > > +.RI "[ device_features " DEVICE_FEATURES "
]"
> > > .RI "[ mac " MACADDR " ]"
> > > .RI "[ mtu " MTU " ]"
> > > .RI "[ max_vqp " MAX_VQ_PAIRS " ]"
> > > @@ -74,6 +75,10 @@ Name of the new vdpa device to add.
> > > Name of the management device to use for device addition.
> > >
> > > .PP
> > > +.BI device_features " DEVICE_FEAETURES"
> >
> > typo
>
> Will fix.
>
> >
> > > +- specifies the device features that is provisioned for the new
vdpa device.
> >
> > I propose
> > the device features -> the virtio "device
features" bit-mask
> >
> > features sounds like it's a generic thing, here's it's
> > the actual binary
> >
> > and maybe add "the bits can be found under
include/uapi/linux/virtio*h,
> > see macros such as VIRTIO_F_ and VIRTIO_NET_F_ for specific bit
values"
>
> That's fine.
>
> >
> > > +This is optional.
> > > +
> >
> > and if not given what are the features?
>
> As in the past, determined by the parent/mgmt device, do we need to
> document this?
I think so, yes.
> >
> > > .BI mac " MACADDR"
> > > - specifies the mac address for the new vdpa device.
> > > This is applicable only for the network type of vdpa device.
This is optional.
> > > @@ -127,6 +132,11 @@ vdpa dev add name foo mgmtdev vdpa_sim_net
> > > Add the vdpa device named foo on the management device
vdpa_sim_net.
> > > .RE
> > > .PP
> > > +vdpa dev add name foo mgmtdev vdpa_sim_net device_features
0x300020000
> > > +.RS 4
> > > +Add the vdpa device named foo on the management device
vdpa_sim_net with device_features of 0x300020000
> > > +.RE
> > > +.PP
> > > vdpa dev add name foo mgmtdev vdpa_sim_net mac 00:11:22:33:44:55
> > > .RS 4
> > > Add the vdpa device named foo on the management device
vdpa_sim_net with mac address of 00:11:22:33:44:55.
> > > diff --git a/vdpa/include/uapi/linux/vdpa.h
b/vdpa/include/uapi/linux/vdpa.h
> > > index 94e4dad1..7c961991 100644
> > > --- a/vdpa/include/uapi/linux/vdpa.h
> > > +++ b/vdpa/include/uapi/linux/vdpa.h
> > > @@ -51,6 +51,7 @@ enum vdpa_attr {
> > > VDPA_ATTR_DEV_QUEUE_INDEX, /* u32 */
> > > VDPA_ATTR_DEV_VENDOR_ATTR_NAME, /* string */
> > > VDPA_ATTR_DEV_VENDOR_ATTR_VALUE, /* u64 */
> > > + VDPA_ATTR_DEV_FEATURES, /* u64 */
> > >
> > > /* new attributes must be added above here */
> > > VDPA_ATTR_MAX,
> > > diff --git a/vdpa/vdpa.c b/vdpa/vdpa.c
> > > index b73e40b4..9a866d61 100644
> > > --- a/vdpa/vdpa.c
> > > +++ b/vdpa/vdpa.c
> > > @@ -27,6 +27,7 @@
> > > #define VDPA_OPT_VDEV_MTU BIT(5)
> > > #define VDPA_OPT_MAX_VQP BIT(6)
> > > #define VDPA_OPT_QUEUE_INDEX BIT(7)
> > > +#define VDPA_OPT_VDEV_FEATURES BIT(8)
> > >
> > > struct vdpa_opts {
> > > uint64_t present; /* flags of present items */
> > > @@ -38,6 +39,7 @@ struct vdpa_opts {
> > > uint16_t mtu;
> > > uint16_t max_vqp;
> > > uint32_t queue_idx;
> > > + __u64 device_features;
> > > };
> > >
> > > struct vdpa {
> >
> > why __u and not uint here?
>
> That's possible, will do.
>
> >
> > > @@ -187,6 +189,17 @@ static int vdpa_argv_u32(struct vdpa *vdpa,
int argc, char **argv,
> > > return get_u32(result, *argv, 10);
> > > }
> > >
> > > +static int vdpa_argv_u64_hex(struct vdpa *vdpa, int argc, char
**argv,
> > > + __u64 *result)
> > > +{
> > > + if (argc <= 0 || !*argv) {
> > > + fprintf(stderr, "number expected\n");
> > > + return -EINVAL;
> > > + }
> > > +
> > > + return get_u64(result, *argv, 16);
> > > +}
> > > +
> > > struct vdpa_args_metadata {
> > > uint64_t o_flag;
> > > const char *err_msg;
> > > @@ -244,6 +257,10 @@ static void vdpa_opts_put(struct nlmsghdr
*nlh, struct vdpa *vdpa)
> > > mnl_attr_put_u16(nlh,
VDPA_ATTR_DEV_NET_CFG_MAX_VQP, opts->max_vqp);
> > > if (opts->present & VDPA_OPT_QUEUE_INDEX)
> > > mnl_attr_put_u32(nlh, VDPA_ATTR_DEV_QUEUE_INDEX,
opts->queue_idx);
> > > + if (opts->present & VDPA_OPT_VDEV_FEATURES) {
> > > + mnl_attr_put_u64(nlh, VDPA_ATTR_DEV_FEATURES,
> > > + opts->device_features);
> > > + }
> > > }
> > >
> > > static int vdpa_argv_parse(struct vdpa *vdpa, int argc, char
**argv,
> > > @@ -329,6 +346,14 @@ static int vdpa_argv_parse(struct vdpa
*vdpa, int argc, char **argv,
> > >
> > > NEXT_ARG_FWD();
> > > o_found |= VDPA_OPT_QUEUE_INDEX;
> > > + } else if (!strcmp(*argv,
"device_features") &&
> > > + (o_optional &
VDPA_OPT_VDEV_FEATURES)) {
> > > + NEXT_ARG_FWD();
> > > + err = vdpa_argv_u64_hex(vdpa, argc, argv,
> > > +
&opts->device_features);
> > > + if (err)
> > > + return err;
> > > + o_found |= VDPA_OPT_VDEV_FEATURES;
> > > } else {
> > > fprintf(stderr, "Unknown option
\"%s\"\n", *argv);
> > > return -EINVAL;
> >
> >
> > should not we validate the value we get? e.g. a mac feature
> > requires that mac is supplied, etc.
>
> This isn't an "issue" that is introduced by this patch.
Management
> device is free to give _F_MAC even if the mac address is not
> provisioned by the userspace. So this should be the responsibility of
> the parent not the netlink/vdpa tool.
right but now user can supply an invalid config. What validates it?
> > in fact hex isn't very user friendly. why not pass feature
> > names instead?
>
> This can be added on top if necessary. In fact there's a plan to
> accept JSON files for provisioning.
>
> The advantages of hex is we don't need to keep the name synced with
> the new features.
>
> Thanks
right but it also means we can't interpret them in any way.
> >
> >
> >
> > > @@ -708,7 +733,7 @@ static int cmd_dev_add(struct vdpa *vdpa, int
argc, char **argv)
> > > err = vdpa_argv_parse_put(nlh, vdpa, argc, argv,
> > > VDPA_OPT_VDEV_MGMTDEV_HANDLE |
VDPA_OPT_VDEV_NAME,
> > > VDPA_OPT_VDEV_MAC |
VDPA_OPT_VDEV_MTU |
> > > - VDPA_OPT_MAX_VQP);
> > > + VDPA_OPT_MAX_VQP |
VDPA_OPT_VDEV_FEATURES);
> > > if (err)
> > > return err;
> > >
> > > --
> > > 2.25.1
> >