diff options
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/bar_init.c | 23 |
1 files changed, 15 insertions, 8 deletions
diff --git a/kernel/bar_init.c b/kernel/bar_init.c index b5703d8..6515613 100644 --- a/kernel/bar_init.c +++ b/kernel/bar_init.c @@ -57,15 +57,22 @@ int nv_vgpu_bar_init(struct nv_vgpu_device *vgpu) if (bar_addr && bar_size) { vgpu->bars[1].gpa = bar_addr; /* - * m143 (1.0 VRAM sizes): expose the EMULATED framebuffer size as - * BAR1's size, not the physical aperture. The guest derives its - * BAR1 from the VFIO region info (config-space BAR writes are - * ignored) and every GSCI/PFB report already carries - * fb_emul_size; leaving the physical 256 MB here made 1G/2G types - * tell the guest one size and expose another (the shadow is - * fb_emul_size, so vidmem bounds and mmap math must match it). + * PROBE (2026-10-08, ISSUE-0001): m143 exposed the emulated + * framebuffer size (512 MB for the turing type) as BAR1; every + * build since the green m140 reference reports BAR1=512MB to + * the guest and the client spins in its first cudaMalloc with + * both counter families written (guest RAM verified). The m140 + * green run reported BAR1=256MB (physical aperture) with the + * same 512 MiB shadow and reached VADD. Restore the physical + * size here as a single-variable probe: if this goes green, + * m143's exposure is the stall trigger and the 1G/2G feature + * must make the guest work with the larger aperture instead of + * assuming it does. (m143: expose the EMULATED framebuffer size + * as BAR1's size, not the physical aperture - the guest derives + * its BAR1 from the VFIO region info and vidmem bounds/mmap + * math must match the shadow.) */ - vgpu->bars[1].size = nv_vgpu_fb_emul_size(vgpu); + vgpu->bars[1].size = bar_size; /* Emulated vidmem: a dedicated host-RAM buffer backing the guest's * VRAM aperture. Do NOT map the host GPU's real BAR1 - guest VRAM * accesses must hit our own clean buffer, not the (broken) host |