Re: [libvirt] [PATCH 2/2] hw/vfio/display: add ramfb support
On Mon, 10 Sep 2018 08:43:40 +0200 Gerd Hoffmann <kraxel@redhat.com> wrote:
So we have a boot display when using a vgpu as primary display.
Use vfio-pci-ramfb instead of vfio-pci to enable it.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com> --- include/hw/vfio/vfio-common.h | 2 ++ hw/vfio/display.c | 12 ++++++++++++ hw/vfio/pci.c | 15 +++++++++++++++ 3 files changed, 29 insertions(+)
diff --git a/include/hw/vfio/vfio-common.h b/include/hw/vfio/vfio-common.h index 821def0565..0d85a0a6f8 100644 --- a/include/hw/vfio/vfio-common.h +++ b/include/hw/vfio/vfio-common.h @@ -26,6 +26,7 @@ #include "qemu/queue.h" #include "qemu/notify.h" #include "ui/console.h" +#include "hw/display/ramfb.h" #ifdef CONFIG_LINUX #include <linux/vfio.h> #endif @@ -146,6 +147,7 @@ typedef struct VFIODMABuf {
typedef struct VFIODisplay { QemuConsole *con; + RAMFBState *ramfb; struct { VFIORegion buffer; DisplaySurface *surface; diff --git a/hw/vfio/display.c b/hw/vfio/display.c index 59c0e5d1d7..3901f580c6 100644 --- a/hw/vfio/display.c +++ b/hw/vfio/display.c @@ -124,6 +124,9 @@ static void vfio_display_dmabuf_update(void *opaque)
primary = vfio_display_get_dmabuf(vdev, DRM_PLANE_TYPE_PRIMARY); if (primary == NULL) { + if (dpy->ramfb) { + ramfb_display_update(dpy->con, dpy->ramfb); + } return; }
@@ -181,6 +184,9 @@ static int vfio_display_dmabuf_init(VFIOPCIDevice *vdev, Error **errp) vdev->dpy->con = graphic_console_init(DEVICE(vdev), 0, &vfio_display_dmabuf_ops, vdev); + if (strcmp(object_get_typename(OBJECT(vdev)), "vfio-pci-ramfb") == 0) { + vdev->dpy->ramfb = ramfb_setup(errp); + } return 0; }
@@ -228,6 +234,9 @@ static void vfio_display_region_update(void *opaque) return; } if (!plane.drm_format || !plane.size) { + if (dpy->ramfb) { + ramfb_display_update(dpy->con, dpy->ramfb); + } return; } format = qemu_drm_format_to_pixman(plane.drm_format); @@ -300,6 +309,9 @@ static int vfio_display_region_init(VFIOPCIDevice *vdev, Error **errp) vdev->dpy->con = graphic_console_init(DEVICE(vdev), 0, &vfio_display_region_ops, vdev); + if (strcmp(object_get_typename(OBJECT(vdev)), "vfio-pci-ramfb") == 0) { + vdev->dpy->ramfb = ramfb_setup(errp); + } return 0; }
diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c index 866f0deeb7..7c0628756e 100644 --- a/hw/vfio/pci.c +++ b/hw/vfio/pci.c @@ -3258,9 +3258,24 @@ static const TypeInfo vfio_pci_dev_info = { }, };
+static void vfio_pci_ramfb_dev_class_init(ObjectClass *klass, void *data) +{ + DeviceClass *dc = DEVICE_CLASS(klass); + + dc->hotpluggable = false; +} + +static const TypeInfo vfio_pci_ramfb_dev_info = { + .name = "vfio-pci-ramfb", + .parent = "vfio-pci", + .instance_size = sizeof(VFIOPCIDevice), + .class_init = vfio_pci_ramfb_dev_class_init, +}; + static void register_vfio_pci_dev_type(void) { type_register_static(&vfio_pci_dev_info); + type_register_static(&vfio_pci_ramfb_dev_info); }
type_init(register_vfio_pci_dev_type)
My concern here is still all of the extra tooling that needs to be added to management layers above QEMU for this device that exists only because we can't hotplug the primary display in QEMU. What happens when we can hotplug the primary display? Aren't disabling hotplug of a vfio-pci device and supporting ramfb two separate things? I think we're leaking current implementation issues out to the device options when really we'd rather have a "ramfb" (or perhaps "console") option on the vfio-pci device and the hotplug capability determined automatically and available through introspection of the device. Thanks, Alex
Hi,
type_register_static(&vfio_pci_dev_info); + type_register_static(&vfio_pci_ramfb_dev_info);
My concern here is still all of the extra tooling that needs to be added to management layers above QEMU for this device that exists only because we can't hotplug the primary display in QEMU. What happens when we can hotplug the primary display?
Ramfb uses fw_cfg, and fw_cfg files can't be added or removed at runtime, the interface simply isn't designed for that.
Aren't disabling hotplug of a vfio-pci device and supporting ramfb two separate things? I think we're leaking current implementation issues out to the device options when really we'd rather have a "ramfb" (or perhaps "console") option on the vfio-pci device and the hotplug capability determined automatically and available through introspection of the device.
Well, I don't think libvirt will have too much trouble handling this. We have two variants (with and without vga compatibility) of other devices: qxl-vga and qxl, virtio-vga and virtio-gpu-pci. libvirt copes just fine and picks the right one (I think depending on video model 'primary' property). Also libvirt manages hotpluggability per device *class*, not per device *instance*. So a device being hotpluggable or not depending on some device property is a problem for libvirt ... I'm open to suggestions how to handle this better, as long as the libvirt people are on board with the approach. cheers, Gerd
On Tue, 11 Sep 2018 06:38:43 +0200 Gerd Hoffmann <kraxel@redhat.com> wrote:
Hi,
type_register_static(&vfio_pci_dev_info); + type_register_static(&vfio_pci_ramfb_dev_info);
My concern here is still all of the extra tooling that needs to be added to management layers above QEMU for this device that exists only because we can't hotplug the primary display in QEMU. What happens when we can hotplug the primary display?
Ramfb uses fw_cfg, and fw_cfg files can't be added or removed at runtime, the interface simply isn't designed for that.
Aren't disabling hotplug of a vfio-pci device and supporting ramfb two separate things? I think we're leaking current implementation issues out to the device options when really we'd rather have a "ramfb" (or perhaps "console") option on the vfio-pci device and the hotplug capability determined automatically and available through introspection of the device.
Well, I don't think libvirt will have too much trouble handling this. We have two variants (with and without vga compatibility) of other devices: qxl-vga and qxl, virtio-vga and virtio-gpu-pci. libvirt copes just fine and picks the right one (I think depending on video model 'primary' property).
Also libvirt manages hotpluggability per device *class*, not per device *instance*. So a device being hotpluggable or not depending on some device property is a problem for libvirt ...
I'm open to suggestions how to handle this better, as long as the libvirt people are on board with the approach.
Ok, so we need a new class to handle making a device non-hotpluggable, but I'm still not sure whether we should make: -device vfio-pci-ramfb or -device vfio-pci-nohotplug,ramfb=on Where ramfb would be a property only available on the nohotplug class variant. The latter seems to provide a lot more flexibility, but which is more practical for libvirt? Thanks, Alex
Hi,
Also libvirt manages hotpluggability per device *class*, not per device *instance*. So a device being hotpluggable or not depending on some device property is a problem for libvirt ...
I'm open to suggestions how to handle this better, as long as the libvirt people are on board with the approach.
Ok, so we need a new class to handle making a device non-hotpluggable, but I'm still not sure whether we should make:
-device vfio-pci-ramfb
or
-device vfio-pci-nohotplug,ramfb=on
Where ramfb would be a property only available on the nohotplug class variant.
I'm fine with the latter.
The latter seems to provide a lot more flexibility, but which is more practical for libvirt?
Any comment from the libvirt camp? thanks, Gerd
On Fri, Sep 14, 2018 at 12:50:09PM +0200, Gerd Hoffmann wrote:
Hi,
Also libvirt manages hotpluggability per device *class*, not per device *instance*. So a device being hotpluggable or not depending on some device property is a problem for libvirt ...
I'm open to suggestions how to handle this better, as long as the libvirt people are on board with the approach.
Ok, so we need a new class to handle making a device non-hotpluggable, but I'm still not sure whether we should make:
-device vfio-pci-ramfb
or
-device vfio-pci-nohotplug,ramfb=on
Where ramfb would be a property only available on the nohotplug class variant.
I'm fine with the latter.
The latter seems to provide a lot more flexibility, but which is more practical for libvirt?
Any comment from the libvirt camp?
We had a discussion about this a few months ago [1] where we spoke about -device vfio-pci-ramfb. However, as Alex has pointed out, the latter proposal gives us more flexibility in terms of introduction of other device properties which are unrelated to ramfb but still might require non-hotpluggable device. Either way, libvirt needs a capability to test whether we should favour this new device over plain vfio-pci if an mdev with display='on' is required. What about new device properties (specifically mdev)? In the discussion below, Gerd noted that apart from the ramfb stuff and the fact that one can be hotplugged while the latter can not, these are identical (option-wise), is that to stay, IOW are we going to keep these two device classes in sync when introducing new vfio-pci device options or are these going to divert more? Is it even possible? What I mean by that is that I'd like to avoid is a situation where there are 2 disjunct sets of options which could potentially lead to problems in decision making in libvirt and we don't like making decisions. Anyhow, I don't feel like any of the proposals has a strong advantage/disadvantage in usage for libvirt, both will require a capability and both would be special cased in our cmdline code depending on the 'display' attribute. Luckily, we don't have mdev migration yet, so it's good we don't have to worry about that at this point yet. [1] https://lists.gnu.org/archive/html/qemu-devel/2018-06/msg03430.html Erik
On Fri, 14 Sep 2018 16:19:07 +0200 Erik Skultety <eskultet@redhat.com> wrote:
On Fri, Sep 14, 2018 at 12:50:09PM +0200, Gerd Hoffmann wrote:
Hi,
Also libvirt manages hotpluggability per device *class*, not per device *instance*. So a device being hotpluggable or not depending on some device property is a problem for libvirt ...
I'm open to suggestions how to handle this better, as long as the libvirt people are on board with the approach.
Ok, so we need a new class to handle making a device non-hotpluggable, but I'm still not sure whether we should make:
-device vfio-pci-ramfb
or
-device vfio-pci-nohotplug,ramfb=on
Where ramfb would be a property only available on the nohotplug class variant.
I'm fine with the latter.
The latter seems to provide a lot more flexibility, but which is more practical for libvirt?
Any comment from the libvirt camp?
We had a discussion about this a few months ago [1] where we spoke about -device vfio-pci-ramfb.
Ah yes, probably my bad for not following up more thoroughly there.
However, as Alex has pointed out, the latter proposal gives us more flexibility in terms of introduction of other device properties which are unrelated to ramfb but still might require non-hotpluggable device. Either way, libvirt needs a capability to test whether we should favour this new device over plain vfio-pci if an mdev with display='on' is required. What about new device properties (specifically mdev)? In the discussion below, Gerd noted that apart from the ramfb stuff and the fact that one can be hotplugged while the latter can not, these are identical (option-wise), is that to stay, IOW are we going to keep these two device classes in sync when introducing new vfio-pci device options or are these going to divert more? Is it even possible? What I mean by that is that I'd like to avoid is a situation where there are 2 disjunct sets of options which could potentially lead to problems in decision making in libvirt and we don't like making decisions.
The vfio-pci device is the parent of this new device, so it should automatically inherit any new properties of vfio-pci, it only modifies the device class for non-hotpluggability and adds properties dependent on non-hotpluggability. I'm not sure if libvirt would expose this as a new model, ie. model="vfio-pci-nohotplug", or if it would be selected via attribute, ie. nohotplug="on", or perhaps if enabling a property only found on the nohotplug variant would select it, ie. ramfb="on". The latter option alone makes it difficult for a user to select it for any random device, for instance if they're trying to setup a kiosk VM where they want to prevent even the guest OS admin from changing the VM configuration. In any case, it seems that libvirt would never be enabling this automatically.
Anyhow, I don't feel like any of the proposals has a strong advantage/disadvantage in usage for libvirt, both will require a capability and both would be special cased in our cmdline code depending on the 'display' attribute. Luckily, we don't have mdev migration yet, so it's good we don't have to worry about that at this point yet.
That's a good point that ramfb depends on display, it seems that regardless of which route we take, using vfio-pci-ramfb or vfio-pci-nohotplug,ramfb=on, it should fail without a display rather than simply adding functionality if a display is present or in the former case, being an obscure way to make a device non-hotpluggable. Personally I prefer the non-hotplug variant of vfio-pci in hopes that it provides more flexibility to users and we only need to tackle this issue once rather than each device we invent with a non-hotplug dependency. Thanks, Alex
On Fri, Sep 14, 2018 at 09:16:33AM -0600, Alex Williamson wrote:
On Fri, 14 Sep 2018 16:19:07 +0200 Erik Skultety <eskultet@redhat.com> wrote:
On Fri, Sep 14, 2018 at 12:50:09PM +0200, Gerd Hoffmann wrote:
Hi,
Also libvirt manages hotpluggability per device *class*, not per device *instance*. So a device being hotpluggable or not depending on some device property is a problem for libvirt ...
I'm open to suggestions how to handle this better, as long as the libvirt people are on board with the approach.
Ok, so we need a new class to handle making a device non-hotpluggable, but I'm still not sure whether we should make:
-device vfio-pci-ramfb
or
-device vfio-pci-nohotplug,ramfb=on
Where ramfb would be a property only available on the nohotplug class variant.
I'm fine with the latter.
The latter seems to provide a lot more flexibility, but which is more practical for libvirt?
Any comment from the libvirt camp?
We had a discussion about this a few months ago [1] where we spoke about -device vfio-pci-ramfb.
Ah yes, probably my bad for not following up more thoroughly there.
However, as Alex has pointed out, the latter proposal gives us more flexibility in terms of introduction of other device properties which are unrelated to ramfb but still might require non-hotpluggable device. Either way, libvirt needs a capability to test whether we should favour this new device over plain vfio-pci if an mdev with display='on' is required. What about new device properties (specifically mdev)? In the discussion below, Gerd noted that apart from the ramfb stuff and the fact that one can be hotplugged while the latter can not, these are identical (option-wise), is that to stay, IOW are we going to keep these two device classes in sync when introducing new vfio-pci device options or are these going to divert more? Is it even possible? What I mean by that is that I'd like to avoid is a situation where there are 2 disjunct sets of options which could potentially lead to problems in decision making in libvirt and we don't like making decisions.
The vfio-pci device is the parent of this new device, so it should automatically inherit any new properties of vfio-pci, it only modifies the device class for non-hotpluggability and adds properties dependent
Okay.
on non-hotpluggability. I'm not sure if libvirt would expose this as a new model, ie. model="vfio-pci-nohotplug", or if it would be selected
although possible, I'd reserve the model attribute to describe the backend API used: vfio-pci, vfio-ap, etc.
via attribute, ie. nohotplug="on", or perhaps if enabling a property
Hmm, I'm afraid there would be a significant amount of confusion around such an attribute, because this would essentially become a feature "sink" + you'd still need to expose the individual properties too, IOW if someone wanted to enable one of x,y,z features they'd need to use 'nohotplug="on"', but in order for libvirt to format the correct cmdline we still need to know that we have to use -device vfio-pci-nohotplug with property ramfb="on" (okay this one can be determined by the existing "display" attribute, but let's look at this in general). I'd go for something more specific, like 'vgaBIOS="on"' so that it immediately does tell more about what it is related to. However, I'm starting to get the feeling that it's not the best idea to stuff all the graphics-related attributes to the master <hostdev> element, but then again, we already kinda started that trend with "display". I'm also wondering whether it wouldn't have been a better idea to have had a video model "vfio-pci" instead of the "none" model, as it would have been more extensible, but then again, <video> describes the device within the guest, not the host...Truth to be told, I'm not really sure what the best approach in libvirt would be, but regardless of what we do in libvirt, it doesn't really matter for QEMU which way you decide to go here as long as QEMU is able to "inform" libvirt about that.
only found on the nohotplug variant would select it, ie. ramfb="on". The latter option alone makes it difficult for a user to select it for any random device, for instance if they're trying to setup a kiosk VM where they want to prevent even the guest OS admin from changing the VM configuration. In any case, it seems that libvirt would never be enabling this automatically.
At first I was playing with the idea that we actually could do that based on the presence of the 'display' attribute, but I guess that would be a very dangerous road to take even though right now I somehow can't see a potential issue, did you think about something specific which we can discuss and therefore find a better solution?
Anyhow, I don't feel like any of the proposals has a strong advantage/disadvantage in usage for libvirt, both will require a capability and both would be special cased in our cmdline code depending on the 'display' attribute. Luckily, we don't have mdev migration yet, so it's good we don't have to worry about that at this point yet.
That's a good point that ramfb depends on display, it seems that regardless of which route we take, using vfio-pci-ramfb or vfio-pci-nohotplug,ramfb=on, it should fail without a display rather than simply adding functionality if a display is present or in the former case, being an obscure way to make a device non-hotpluggable. Personally I prefer the non-hotplug variant of vfio-pci in hopes that it provides more flexibility to users and we only need to tackle this issue once rather than each device we invent with a non-hotplug dependency. Thanks,
Alex
Erik
participants (3)
-
Alex Williamson -
Erik Skultety -
Gerd Hoffmann