Hi, On Wed, Nov 11, 2026 at 12:13:16PM +0100, [email protected] wrote: > Hi, > > I am a Windows developer porting some of my own desktop tools to Linux, and > I use a Debian guest on Hyper-V to judge how they look and feel there. > > The static picture is good, the mouse pointer is not, and I think the cause > is small and fixable in hyperv_drm. I am writing because this is the first > thing a Windows developer notices when trying Linux on their own machine: as > long as a Linux VM feels sluggish next to a Windows VM on the same host, few > of us will move a repo over and be happy about it. > > Setup > - Host: [Windows 11/26H2], Hyper-V, Generation 2 VM, basic session > (vmconnect), 3840x2160 display > - Guest: Debian 13, kernel 6.12.113+deb13-amd64, hyperv_drm (Synthvid 3.5), > KDE Plasma 6.3.6 on Wayland (KWin 6.3.6, libwayland 1.23.1) > - 4 vCPUs, about 7 GB RAM > > What I saw so far: > > - The pointer trails the hand by a small constant delay. Fast, long > movements (taskbar to top of the screen) are visibly laggy, not just > slightly off. > > - A Windows guest on the same host, same viewer, same mouse is smooth - > directly, and even when the host itself is reached through RDP. The Debian > guest lags in both cases. > > - While moving the pointer: kwin_wayland uses 35-50 % of one core, a kworker > (events_unbound) 13-21 %, and the VM is about 80 % idle. Adding vCPUs > changes nothing, so it is not a lack of CPU. > > - A smaller cursor and disabling KDE's shake-to-enlarge effect helped a > little. A plain color desktop did not. > > What I think the cause is (from reading the driver, I have not patched it) > > 1. hyperv_drm exposes only the primary plane. There is no cursor plane, so > the compositor has to paint the pointer into the framebuffer, and every > movement goes through compositing, the "blit to VRAM" and the dirt message. > > 2. hyperv_pipe_update() uses drm_atomic_helper_damage_merged(), so the old > and the new pointer position become one bounding rectangle. For a long or > diagonal move at 4K that is a large part of the screen per pointer step, > which matches the more noticable lag (much worse) on big movements. > > 3. The protocol already has what is needed. hyperv_hide_hw_ptr() exists > because, as its commit message says, "Hyper-V supports a hardware cursor > feature. It is not used by Linux VM, but the Hyper-V host still draws" a > pointer; the driver hides it by sending a 1x1 transparent ARGB shape with > SYNTHVID_POINTER_SHAPE. So the host does accept a shape from a Linux guest > and does draw the pointer itself. > > Proposal: > > Add a cursor plane to hyperv_drm (up to SYNTHVID_CURSOR_MAX_X/Y, 96x96, > ARGB) that sends the compositor's cursor image and hot spot through > SYNTHVID_POINTER_SHAPE instead of the transparent pixel, and re-sends it on > SYNTHVID_FEATURE_CHANGE where the hide call is made today. >
Cursor plane support is a WIP, you can test out [1] on your end. [1] https://patchwork.freedesktop.org/patch/761562/ Hamza > With cursor hot-spot support for virtualized drivers, compositors would then > leave the pointer to the host, as Windows guests do, and pointer movement > would cause no framebuffer damage at all. > > Independently of that, handling damage clips individually instead of merged > would reduce the cost of the software-cursor fallback (cursors larger than > 96x96). > > A second, smaller item found on the way: > > With Set-VMVideo set to 3840x2160 (Maximum), the guest offers modes only up > to 2560x1600. The connector's modes come from drm_add_modes_noedid(), and > 3840x2160 is not in that table, so it only appears with > video=Virtual-1:3840x2160 on the kernel command line. Adding the host's > maximum (or preferred) resolution as an explicit mode would make 4K hosts > work without that step. > > I can test patches on this setup and report back. I am not in a position to > write the kernel change myself (drivers are not my field of work at all). > > Thanks for maintaining the driver. > > Sascha Eder > >
