On Fri, Jan 19, 2024 at 4:22 AM Thomas Zimmermann <tzimmermann@xxxxxxx> wrote: > > Hi > > Am 18.01.24 um 19:25 schrieb Zack Rusin: > > On Mon, Jan 15, 2024 at 3:21 AM Thomas Zimmermann <tzimmermann@xxxxxxx> wrote: > >> > >> Hi > >> > >> Am 12.01.24 um 21:38 schrieb Ian Forbes: > >>> SVGA requires surfaces to fit within graphics memory (max_mob_pages) which > >>> means that modes with a final buffer size that would exceed graphics memory > >>> must be pruned otherwise creation will fail. > >>> > >>> Additionally, device commands which use multiple graphics resources must > >>> have all their resources fit within graphics memory for the duration of the > >>> command. Thus we need a small carve out of 1/4 of graphics memory to ensure > >>> commands likes surface copies to the primary framebuffer for cursor > >>> composition or damage clips can fit within graphics memory. > >>> > >>> This fixes an issue where VMs with low graphics memory (< 64MiB) configured > >>> with high resolution mode boot to a black screen because surface creation > >>> fails. > >> > >> That is a long-standing problem, which we have observed with other > >> drivers as well. On low-memory devices, TTM doesn't play well. The real > >> fix would be to export all modes that possibly fit and sort out the > >> invalid configurations in atomic_check. It's just a lot more work. > >> > >> Did you consider simply ignoring vmwgfx devices with less than 64 MiB of > >> VRAM? > > > > Unfortunately we can't do that because on new esx servers without > > gpu's the default is 16MB. A lot of people are still running their esx > > boxes with 4MB, which is in general the most common problem because > > with 4MB people still tend to like to set 1280x800 which with 32bpp fb > > takes 4096000 bytes and with 4MB available that leaves only 96KB > > available and we need more to also allocate things like the cursor. > > Even if ttm did everything right technically 1280x800 @ 32bpp > > resolution will fit in a 4MB graphics memory, but then the system will > > not be able to have a hardware (well, virtualized) cursor. It's > > extremely unlikely people would even be aware of this tradeoff when > > making the decision to increase resolution. > > Do you allocate buffer storage directly in the provided VRAM? If so how > do you do page flips then? You'd need for the example of 1280x800-32, > you'd need around 8 MiB to keep front and back buffer in VRAM. I guess, > you only support the framebuffer console (which doesn't do pageflips)? In general, yes. Of course it's a little more convoluted because we'll act like OpenGL runtime here (i.e. glXSwapBuffers), i.e. our driver will fake page-flips because the only memory we'll have is a single buffer as the actual page-flipping happens in the presentation code on the host. So the guest is not aware of the actual presentation (it's also why we don't have any sort of vblank signaling in vmwgfx, the concept just doesn't exist for us). i.e. on para-virtualized drivers the actual page-flips will be property of the presentation code that's outside of the guest. It's definitely one those things that I wanted to have a good solution for in a while, in particular to have a better story behind vblank handling, but it's difficult because "presentation" on vm's is in general difficult to define - it might be some vnc connected host on the other continent. Having said that that's basically a wonky VRR display so we should be able to handle our presentation as VRR and give more control of updates to the guest, but we haven't done it yet. > In mgag200 and ast, I had the luxury for replacing TTM with SHMEM > helpers, which worked around the problem easily. Maybe that's an option > for low-memory systems? Our current device doesn't have the ability to present out of unspecified memory in the guest, i.e. the host, which is doing the presentation, is not aware of how guest kernel lays out the memory so we need to basically create a page-table for every graphics object (VMW_PL_MOB placement in vmwgfx) so that the host can actually find the memory it needs to read. So the shmem helpers would need something extra for us to be able to generate those page tables for the drm_gem_object's it deals with. z
