drm/nouveau/uvmm: Allow larger pages

Now that everything in UVMM knows about the variable page shift, we can
select larger values.

The proposed approach relies on nouveau_bo::page unless if it would cause
alignment issues (in which case we fall back to searching for an
appropriate shift)

Signed-off-by: Mary Guillemard <mary@mary.zone>
Co-developed-by: Mohamed Ahmed <mohamedahmedegypt2001@gmail.com>
Signed-off-by: Mohamed Ahmed <mohamedahmedegypt2001@gmail.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Reviewed-by: James Jones <jajones@nvidia.com>
Signed-off-by: Lyude Paul <lyude@redhat.com>
Link: https://patch.msgid.link/20251110-nouveau-compv6-v6-2-83b05475f57c@mary.zone

authored by

Mary Guillemard and committed by

Lyude Paul 7 months ago c488a94e 6d134953

+58 -2

1 changed file

expand all

drivers

gpu

drm

nouveau

nouveau_uvmm.c

+58 -2

drivers/gpu/drm/nouveau/nouveau_uvmm.c

··· 454 454 drm_gpuva_insert(va->vm, va); 455 455 } 456 456 457 + static bool 458 + op_map_aligned_to_page_shift(const struct drm_gpuva_op_map *op, u8 page_shift) 459 + { 460 + u64 non_page_bits = (1ULL << page_shift) - 1; 461 + 462 + return (op->va.addr & non_page_bits) == 0 && 463 + (op->va.range & non_page_bits) == 0 && 464 + (op->gem.offset & non_page_bits) == 0; 465 + } 466 + 467 + static u8 468 + select_page_shift(struct nouveau_uvmm *uvmm, struct drm_gpuva_op_map *op) 469 + { 470 + struct nouveau_bo *nvbo = nouveau_gem_object(op->gem.obj); 471 + 472 + /* nouveau_bo_fixup_align() guarantees that the page size will be aligned 473 + * for most cases, but it can't handle cases where userspace allocates with 474 + * a size and then binds with a smaller granularity. So in order to avoid 475 + * breaking old userspace, we need to ensure that the VA is actually 476 + * aligned before using it, and if it isn't, then we downgrade to the first 477 + * granularity that will fit, which is optimal from a correctness and 478 + * performance perspective. 479 + */ 480 + if (op_map_aligned_to_page_shift(op, nvbo->page)) 481 + return nvbo->page; 482 + 483 + struct nouveau_mem *mem = nouveau_mem(nvbo->bo.resource); 484 + struct nvif_vmm *vmm = &uvmm->vmm.vmm; 485 + int i; 486 + 487 + /* If the given granularity doesn't fit, let's find one that will fit. */ 488 + for (i = 0; i < vmm->page_nr; i++) { 489 + /* Ignore anything that is bigger or identical to the BO preference. */ 490 + if (vmm->page[i].shift >= nvbo->page) 491 + continue; 492 + 493 + /* Skip incompatible domains. */ 494 + if ((mem->mem.type & NVIF_MEM_VRAM) && !vmm->page[i].vram) 495 + continue; 496 + if ((mem->mem.type & NVIF_MEM_HOST) && 497 + (!vmm->page[i].host || vmm->page[i].shift > PAGE_SHIFT)) 498 + continue; 499 + 500 + /* If it fits, return the proposed shift. */ 501 + if (op_map_aligned_to_page_shift(op, vmm->page[i].shift)) 502 + return vmm->page[i].shift; 503 + } 504 + 505 + /* If we get here then nothing can reconcile the requirements. This should never 506 + * happen. 507 + */ 508 + drm_WARN_ONCE(op->gem.obj->dev, 1, "Could not find an appropriate page size.\n"); 509 + 510 + return PAGE_SHIFT; 511 + } 512 + 457 513 static void 458 514 nouveau_uvmm_sm_prepare_unwind(struct nouveau_uvmm *uvmm, 459 515 struct nouveau_uvma_prealloc *new, ··· 562 506 if (vmm_get_range) 563 507 nouveau_uvmm_vmm_put(uvmm, vmm_get_start, 564 508 vmm_get_range, 565 - PAGE_SHIFT); 509 + select_page_shift(uvmm, &op->map)); 566 510 break; 567 511 } 568 512 case DRM_GPUVA_OP_REMAP: { ··· 655 599 656 600 uvma->region = args->region; 657 601 uvma->kind = args->kind; 658 - uvma->page_shift = PAGE_SHIFT; 602 + uvma->page_shift = select_page_shift(uvmm, op); 659 603 660 604 drm_gpuva_map(&uvmm->base, &uvma->va, op); 661 605

Configure Feed

Configure Feed