From 31594b263cc4fbe30ebbe412a4a788f726f03a02 Mon Sep 17 00:00:00 2001 From: Arthur Hanson Rasmusson Date: Sun, 4 Oct 2026 16:19:31 -0400 Subject: vgsp: end-to-end CUDA acceptance in the guest The no-claim emulation path now runs a stock CUDA program end to end: device properties, cudaMalloc, a GPU-authored string kernel read back byte-for-byte, two 4 MB host-to-device copies, a 1 M-element vectorAdd executed on the physical GPU through the host bridge, and the device-to-host copy-back verified element by element. Three consecutive supervised cycles report MAIL + VADD bad=0 checksum=392344090.500000 + STATUS OK with identical input/output hashes. Two data-path correctness fixes make the transfers sound: - Once-only op execution per GPFIFO entry. The pump replays recently consumed ring entries when translation batches are injected so missed releases can land; re-executing a CE copy is not idempotent because its source is a client staging buffer that may have been refilled. Slots now record their raw entry words and a replayed entry matching that identity suppresses operations while still processing releases. - Per-page translation of virtual operands. svc_copy resolved only the first page of a VA range and walked the PA linearly; client staging ranges are frequently split (observed 195 + 61 pages), so the tail of alternating chunks was read from unrelated memory. Multi-page copies now re-resolve each 4 KB page through the trusted resolver. Also carries the earlier uncommitted emulation fixes that got the client to the launch (QMD address-window gate, counter registration persistence, 512 MiB emulated FB, bridge slot hygiene) and bounded diagnostics (sema-bigop, sema-bigcopy, bridge-in/out hashes, daemon INCHECK/OUTCHECK/dumps). --- kernel/include/nvidia-vgpu-vfio.h | 20 +++++++++++++++++++- 1 file changed, 19 insertions(+), 1 deletion(-) (limited to 'kernel/include/nvidia-vgpu-vfio.h') diff --git a/kernel/include/nvidia-vgpu-vfio.h b/kernel/include/nvidia-vgpu-vfio.h index 1c2f67b..ccd7262 100644 --- a/kernel/include/nvidia-vgpu-vfio.h +++ b/kernel/include/nvidia-vgpu-vfio.h @@ -25,7 +25,18 @@ #define NV_VGPU_MAX_IRQS NVIDIA_VGPU_MAX_IRQS #define NV_VGPU_BAR0_SIZE (16 * 1024 * 1024) -#define NV_VGPU_BAR1_SIZE (256 * 1024 * 1024) +/* + * BAR1 (VRAM aperture) and the emulated framebuffer size. These two move + * together: the guest derives its BAR1 size from fbLength + * (kbusInitBarsSize_VF: pciBarSizes[1] = pGpu->fbLength), and its vidmem + * address space is [0, fbLength). The shadow backing bars[1].vaddr is + * NV_VGPU_BAR1_SIZE, so any guest vidmem PA the emulation translates must + * stay inside it. 512 MiB gives UVM enough headroom after its own + * reservations to satisfy the acceptance program's 3x4 MiB vadd buffers + * (at 256 MiB the guest reported only ~5 MiB free). + */ +#define NV_VGPU_BAR1_SIZE (512 * 1024 * 1024) +#define NV_VGPU_EMUL_FB_SIZE (512ULL * 1024 * 1024) #define NV_VGPU_BAR2_SIZE (32 * 1024 * 1024) #define NV_VGPU_BAR3_SIZE (32 * 1024 * 1024) @@ -56,6 +67,13 @@ struct nv_vgpu_bar { void *vaddr; bool mapped; bool trapped; + /* + * True when vaddr is a host-RAM shadow (vzalloc) owned by this module + * instead of an ioremap'd physical BAR. Guest mappings must then remap + * the shadow's own pages; a real BAR must never be exposed to the + * guest outside the supervised physical claim. + */ + bool emulated; struct mutex lock; }; -- cgit v1.2.3