Fixing nouveau for >4k PAGE_SIZE
Benjamin Herrenschmidt
benh at kernel.crashing.org
Sun Aug 11 02:35:13 PDT 2013
On Sun, 2013-08-11 at 11:02 +0200, Maarten Lankhorst wrote:
> > diff --git a/drivers/gpu/drm/nouveau/core/engine/fifo/nv40.c b/drivers/gpu/drm/nouveau/core/engine/fifo/nv40.c
> > index 5c7433d..c314a5f 100644
> > --- a/drivers/gpu/drm/nouveau/core/engine/fifo/nv40.c
> > +++ b/drivers/gpu/drm/nouveau/core/engine/fifo/nv40.c
> > @@ -190,8 +190,8 @@ nv40_fifo_chan_ctor(struct nouveau_object *parent,
> > if (size < sizeof(*args))
> > return -EINVAL;
> >
> > - ret = nouveau_fifo_channel_create(parent, engine, oclass, 0, 0xc00000,
> > - 0x1000, args->pushbuf,
> > + ret = nouveau_fifo_channel_create(parent, engine, oclass, 0, 0x800000,
> > + 0x10000, args->pushbuf,
> > (1ULL << NVDEV_ENGINE_DMAOBJ) |
> > (1ULL << NVDEV_ENGINE_SW) |
> > (1ULL << NVDEV_ENGINE_GR) |
> Why the size change?
This reverts the value to older ones, however that patch might not be
needed anymore (I was carrying it from Dave but if we don't map the
registers into userspace we shouldn't need to force align them).
> > diff --git a/drivers/gpu/drm/nouveau/core/subdev/vm/base.c b/drivers/gpu/drm/nouveau/core/subdev/vm/base.c
> > index ef3133e..5833851 100644
> > --- a/drivers/gpu/drm/nouveau/core/subdev/vm/base.c
> > +++ b/drivers/gpu/drm/nouveau/core/subdev/vm/base.c
> > @@ -84,10 +84,11 @@ nouveau_vm_map_sg_table(struct nouveau_vma *vma, u64 delta, u64 length,
> > {
> > struct nouveau_vm *vm = vma->vm;
> > struct nouveau_vmmgr *vmm = vm->vmm;
> > - int big = vma->node->type != vmm->spg_shift;
> > + u32 shift = vma->node->type;
> > + int big = shift != vmm->spg_shift;
> > u32 offset = vma->node->offset + (delta >> 12);
> > - u32 bits = vma->node->type - 12;
> > - u32 num = length >> vma->node->type;
> > + u32 bits = shift - 12;
> > + u32 num = length >> shift;
> > u32 pde = (offset >> vmm->pgt_bits) - vm->fpde;
> > u32 pte = (offset & ((1 << vmm->pgt_bits) - 1)) >> bits;
> > u32 max = 1 << (vmm->pgt_bits - bits);
> > @@ -98,7 +99,7 @@ nouveau_vm_map_sg_table(struct nouveau_vma *vma, u64 delta, u64 length,
> >
> > for_each_sg(mem->sg->sgl, sg, mem->sg->nents, i) {
> > struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
> > - sglen = sg_dma_len(sg) >> PAGE_SHIFT;
> > + sglen = sg_dma_len(sg) >> shift;
> >
> > end = pte + sglen;
> > if (unlikely(end >= max))
> Please add a WARN_ON(big); in map_sg and map_sg_table if you do this.
So that's debatable :-)
The above code is map_sg. Arguably my patch *fixes* using it with card
large pages :-)
IE, Look at the "usual" case (PAGE_SHIFT=12). Today, the above means
sglen will be in units of small pages. But everything else in that loop
operates in units of whatever is represented by the pte, which can
either be 4k or larger.
So adding "sglen" to "pte" was never right for shift != 12 before
(regardless of PAGE_SHIFT).
With my patch, it's more correct, so as such, adding a WARN_ON here
wouldn't be "if I do this" :-)
Now, the "big" case still cannot really work here with PAGE_SHIFT=12,
because that would require the sg segments to be multiple of
"shift" (the card large page) and that is not going to be the case.
So funnily enough, we *could* use card large pages of 64k ("big") if ...
we had PAGE_SHIFT=16 ... which we do on ppc64 :-) But not anywhere else.
But yes, the point remains, in the general case, that function cannot
work for the "big" case, so I wonder if we should catch it with a
WARN_ON and maybe simplify the code a bunch while at it...
> > @@ -106,7 +107,7 @@ nouveau_vm_map_sg_table(struct nouveau_vma *vma, u64 delta, u64 length,
> > len = end - pte;
> >
> > for (m = 0; m < len; m++) {
> > - dma_addr_t addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
> > + dma_addr_t addr = sg_dma_address(sg) + (m << shift);
> >
> > vmm->map_sg(vma, pgt, mem, pte, 1, &addr);
> > num--;
> > @@ -121,7 +122,7 @@ nouveau_vm_map_sg_table(struct nouveau_vma *vma, u64 delta, u64 length,
> > }
> > if (m < sglen) {
> > for (; m < sglen; m++) {
> > - dma_addr_t addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
> > + dma_addr_t addr = sg_dma_address(sg) + (m << shift);
> >
> > vmm->map_sg(vma, pgt, mem, pte, 1, &addr);
> > num--;
> > @@ -136,6 +137,7 @@ finish:
> > vmm->flush(vm);
> > }
> >
> > +#if PAGE_SHIFT == 12
> > void
> > nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
> > struct nouveau_mem *mem)
> > @@ -143,10 +145,11 @@ nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
> > struct nouveau_vm *vm = vma->vm;
> > struct nouveau_vmmgr *vmm = vm->vmm;
> > dma_addr_t *list = mem->pages;
> > - int big = vma->node->type != vmm->spg_shift;
> > + u32 shift = vma->node->type;
> > + int big = shift != vmm->spg_shift;
> > u32 offset = vma->node->offset + (delta >> 12);
> > - u32 bits = vma->node->type - 12;
> > - u32 num = length >> vma->node->type;
> > + u32 bits = shift - 12;
> > + u32 num = length >> shift;
> > u32 pde = (offset >> vmm->pgt_bits) - vm->fpde;
> > u32 pte = (offset & ((1 << vmm->pgt_bits) - 1)) >> bits;
> > u32 max = 1 << (vmm->pgt_bits - bits);
> > @@ -173,6 +176,52 @@ nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
So the above is the existing one which I kept mostly intact ... but
cannot work for shift != 12 either for the same reasons so here too, if
we could simplify that ...
> > vmm->flush(vm);
> > }
> > +#else
> > +void
> > +nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
> > + struct nouveau_mem *mem)
> > +{
> > + struct nouveau_vm *vm = vma->vm;
> > + struct nouveau_vmmgr *vmm = vm->vmm;
> > + dma_addr_t *list = mem->pages;
> > + u32 shift = vma->node->type;
> > + int big = shift != vmm->spg_shift;
> > + u32 offset = vma->node->offset + (delta >> 12);
> > + u32 bits = shift - 12;
> > + u32 num = length >> shift;
> > + u32 pde = (offset >> vmm->pgt_bits) - vm->fpde;
> > + u32 pte = (offset & ((1 << vmm->pgt_bits) - 1)) >> bits;
> > + u32 max = 1 << (vmm->pgt_bits - bits);
> > + u32 sub_cnt = 1 << (PAGE_SHIFT - shift);
> > + u32 sub_rem = 0;
> > + u64 phys = 0;
> > +
> > +
> > + /* XXX This will not work for a big mapping ! */
> > + WARN_ON_ONCE(big);
> > +
> > + while (num) {
> > + struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
> > +
> > + if (sub_rem == 0) {
> > + phys = *(list++);
> > + sub_rem = sub_cnt;
> > + }
> > + vmm->map_sg(vma, pgt, mem, pte, 1, &phys);
> > +
> > + num -= 1;
> > + pte += 1;
> > + sub_rem -= 1;
> > + phys += 1 << shift;
> > + if (unlikely(pte >= max)) {
> > + pde++;
> > + pte = 0;
> > + }
> > + }
> > +
> > + vmm->flush(vm);
> > +}
> > +#endif
>
> Considering that map_sg in nv04-50 takes PAGE_SIZE pages, it would be easier to fix
> map_sg for nv50 and nvc0, this would mean less fixing in map_sg/map_sg_table.
Sorry, I'm not sure what you mean exactly ... The code *today* tries to
handle things at the low level (vmm->map_sg) and that cannot work the
way it's done. I'm removing that. Or rather, that cannot work unless we
can guarantee that there will be no crossing of PTE page boundary (no
crossing of pde) by the PAGE_SIZE page, which I think we cannot (see my
discussion in my email asking if we could enforce that somewhat).
Additionally the current code is broken in that the upper layer in
vm/base.c doesn't increment "pte" by the right amount.
Now, if those two assertions can be made always true:
- Those two functions (map_sg and map_sg_table) never deal with the
"big" case.
- An object is always mapped at a card address that is a multiple
of PAGE_SIZE (ie, invividual PAGE_SIZE pages don't cross pde boundaries
when mapped)
Then we can probably simplify the code significantly *and* go back to
handling PAGE_SIZE pages in the vmm->map_sg() instead of breaking them
up a level above like I do.
> I don't like the duplicate definition, and the extra for loop in map_sg will be compiled out when page_size == 12.
Sort-of. Right now, my code prevents calling vmm->map_sg with more than
"1" as len which reduces perf in the PAGE_SHIFT=12 case, which is why I
did the duplication.
However this is just an illustration. I'm fully aware that this is not
good as-is, I'm just poking to see what you guys think is the best
approach to *properly* fix it.
> Oh fun fact:
> on nv50 if PAGE_SIZE = 64K you should just use large pages on the nvidia card for everything. :D
I don't think I care at this stage but heh ...
> I have no idea if it works for bar, but you could test..
> In subdev/bar/nv50.c kmap/umap change the 12 argument to 16, and change vm->pgt[0].obj[0] to vm->pgt[0].obj[1] and vm->pgt[0].refcount[0] to vm->pgt[0].refcount[1].
>
> for nvc0 that would require 128K pages, but I believe it should work too.
I don't think we want to go there right now, there is little benefit for
the vast majority of platforms (x86 and ARM). Let's stick to making
ppc64 "work" and not bother too much with things that will never be used
in practice :-)
I'd rather make the code simpler by removing the "big" case from those
functions if it's never going to be used.
Cheers,
Ben.
> >
> > void
> > nouveau_vm_unmap_at(struct nouveau_vma *vma, u64 delta, u64 length)
> > diff --git a/drivers/gpu/drm/nouveau/core/subdev/vm/nv04.c b/drivers/gpu/drm/nouveau/core/subdev/vm/nv04.c
> > index ed45437..f7e2311 100644
> > --- a/drivers/gpu/drm/nouveau/core/subdev/vm/nv04.c
> > +++ b/drivers/gpu/drm/nouveau/core/subdev/vm/nv04.c
> > @@ -39,14 +39,10 @@ nv04_vm_map_sg(struct nouveau_vma *vma, struct nouveau_gpuobj *pgt,
> > {
> > pte = 0x00008 + (pte * 4);
> > while (cnt) {
> > - u32 page = PAGE_SIZE / NV04_PDMA_PAGE;
> > u32 phys = (u32)*list++;
> > - while (cnt && page--) {
> > - nv_wo32(pgt, pte, phys | 3);
> > - phys += NV04_PDMA_PAGE;
> > - pte += 4;
> > - cnt -= 1;
> > - }
> > + nv_wo32(pgt, pte, phys | 3);
> > + pte += 4;
> > + cnt -= 1;
> > }
> > }
> >
> > diff --git a/drivers/gpu/drm/nouveau/core/subdev/vm/nv41.c b/drivers/gpu/drm/nouveau/core/subdev/vm/nv41.c
> > index 064c762..a78f624 100644
> > --- a/drivers/gpu/drm/nouveau/core/subdev/vm/nv41.c
> > +++ b/drivers/gpu/drm/nouveau/core/subdev/vm/nv41.c
> > @@ -43,14 +43,10 @@ nv41_vm_map_sg(struct nouveau_vma *vma, struct nouveau_gpuobj *pgt,
> > {
> > pte = pte * 4;
> > while (cnt) {
> > - u32 page = PAGE_SIZE / NV41_GART_PAGE;
> > u64 phys = (u64)*list++;
> > - while (cnt && page--) {
> > - nv_wo32(pgt, pte, (phys >> 7) | 1);
> > - phys += NV41_GART_PAGE;
> > - pte += 4;
> > - cnt -= 1;
> > - }
> > + nv_wo32(pgt, pte, (phys >> 7) | 1);
> > + pte += 4;
> > + cnt -= 1;
> > }
> > }
> See above^, it's better if you could fixup nv50/nvc0.c instead.
> >
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
> > index af20fba..694024d 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_bo.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
> > @@ -226,7 +226,7 @@ nouveau_bo_new(struct drm_device *dev, int size, int align,
> > nvbo->page_shift = 12;
> > if (drm->client.base.vm) {
> > if (!(flags & TTM_PL_FLAG_TT) && size > 256 * 1024)
> > - nvbo->page_shift = drm->client.base.vm->vmm->lpg_shift;
> > + nvbo->page_shift = lpg_shift;
> > }
> >
> > nouveau_bo_fixup_align(nvbo, flags, &align, &size);
> Hm.. I don't know if it will be an issue here. But I'm concerned about the cases where nouveau_vm can end up unaligned.
> This will not be an issue for the bar mappings, because they map the entire bo, and bo's are always page aligned.
> See nouveau_bo_fixup_align.
>
> But the channel vm mappings are no longer guaranteed to be page aligned. In fact it's very likely they are all misaligned due to some vm allocations
> done when mapping a channel. This is only a problem on nv50+ though. Probably not an issue you're hitting.
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
> > index ca5492a..494cf88 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
> > @@ -31,7 +31,7 @@ nv04_sgdma_bind(struct ttm_tt *ttm, struct ttm_mem_reg *mem)
> > {
> > struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)ttm;
> > struct nouveau_mem *node = mem->mm_node;
> > - u64 size = mem->num_pages << 12;
> > + u64 size = mem->num_pages << PAGE_SHIFT;
> >
> > if (ttm->sg) {
> > node->sg = ttm->sg;
> > diff --git a/drivers/gpu/drm/nouveau/nouveau_ttm.c b/drivers/gpu/drm/nouveau/nouveau_ttm.c
> > index 19e3757..f0629de 100644
> > --- a/drivers/gpu/drm/nouveau/nouveau_ttm.c
> > +++ b/drivers/gpu/drm/nouveau/nouveau_ttm.c
> > @@ -252,8 +252,8 @@ nv04_gart_manager_new(struct ttm_mem_type_manager *man,
> >
> > node->page_shift = 12;
> >
> > - ret = nouveau_vm_get(man->priv, mem->num_pages << 12, node->page_shift,
> > - NV_MEM_ACCESS_RW, &node->vma[0]);
> > + ret = nouveau_vm_get(man->priv, mem->num_pages << PAGE_SHIFT,
> > + node->page_shift, NV_MEM_ACCESS_RW, &node->vma[0]);
> > if (ret) {
> > kfree(node);
> > return ret;
> >
> >
More information about the dri-devel
mailing list