#include "drmP.h"
#include "drm.h"
+
#include "nouveau_drv.h"
+#include "nouveau_vm.h"
+
+#define BAR1_VM_BASE 0x0020000000ULL
+#define BAR1_VM_SIZE pci_resource_len(dev->pdev, 1)
+#define BAR3_VM_BASE 0x0000000000ULL
+#define BAR3_VM_SIZE pci_resource_len(dev->pdev, 3)
struct nv50_instmem_priv {
uint32_t save1700[5]; /* 0x1700->0x1710 */
- struct nouveau_gpuobj *pramin_pt;
- struct nouveau_gpuobj *pramin_bar;
- struct nouveau_gpuobj *fb_bar;
+ struct nouveau_gpuobj *bar1_dmaobj;
+ struct nouveau_gpuobj *bar3_dmaobj;
};
static void
return;
nouveau_gpuobj_ref(NULL, &chan->ramfc);
+ nouveau_vm_ref(NULL, &chan->vm, chan->vm_pd);
nouveau_gpuobj_ref(NULL, &chan->vm_pd);
if (chan->ramin_heap.free_stack.next)
drm_mm_takedown(&chan->ramin_heap);
}
static int
-nv50_channel_new(struct drm_device *dev, u32 size,
+nv50_channel_new(struct drm_device *dev, u32 size, struct nouveau_vm *vm,
struct nouveau_channel **pchan)
{
struct drm_nouveau_private *dev_priv = dev->dev_private;
u32 pgd = (dev_priv->chipset == 0x50) ? 0x1400 : 0x0200;
u32 fc = (dev_priv->chipset == 0x50) ? 0x0000 : 0x4200;
struct nouveau_channel *chan;
- int ret;
+ int ret, i;
chan = kzalloc(sizeof(*chan), GFP_KERNEL);
if (!chan)
return ret;
}
+ for (i = 0; i < 0x4000; i += 8) {
+ nv_wo32(chan->vm_pd, i + 0, 0x00000000);
+ nv_wo32(chan->vm_pd, i + 4, 0xdeadcafe);
+ }
+
+ ret = nouveau_vm_ref(vm, &chan->vm, chan->vm_pd);
+ if (ret) {
+ nv50_channel_del(&chan);
+ return ret;
+ }
+
ret = nouveau_gpuobj_new_fake(dev, chan->ramin->pinst == ~0 ? ~0 :
chan->ramin->pinst + fc,
chan->ramin->vinst + fc, 0x100,
struct drm_nouveau_private *dev_priv = dev->dev_private;
struct nv50_instmem_priv *priv;
struct nouveau_channel *chan;
+ struct nouveau_vm *vm;
int ret, i;
+ u64 nongart_o;
u32 tmp;
priv = kzalloc(sizeof(*priv), GFP_KERNEL);
ret = drm_mm_init(&dev_priv->ramin_heap, 0, dev_priv->ramin_size);
if (ret) {
NV_ERROR(dev, "Failed to init RAMIN heap\n");
- return -ENOMEM;
+ goto error;
}
- /* we need a channel to plug into the hw to control the BARs */
- ret = nv50_channel_new(dev, 128*1024, &dev_priv->fifos[0]);
+ /* BAR3 */
+ ret = nouveau_vm_new(dev, BAR3_VM_BASE, BAR3_VM_SIZE, BAR3_VM_BASE,
+ 29, 12, 16, &dev_priv->bar3_vm);
if (ret)
- return ret;
- chan = dev_priv->fifos[127] = dev_priv->fifos[0];
+ goto error;
- /* allocate page table for PRAMIN BAR */
- ret = nouveau_gpuobj_new(dev, chan, (dev_priv->ramin_size >> 12) * 8,
- 0x1000, NVOBJ_FLAG_ZERO_ALLOC,
- &priv->pramin_pt);
+ ret = nouveau_gpuobj_new(dev, NULL, (BAR3_VM_SIZE >> 12) * 8,
+ 0x1000, NVOBJ_FLAG_DONT_MAP |
+ NVOBJ_FLAG_ZERO_ALLOC,
+ &dev_priv->bar3_vm->pgt[0].obj);
if (ret)
- return ret;
+ goto error;
+ dev_priv->bar3_vm->pgt[0].page_shift = 12;
+ dev_priv->bar3_vm->pgt[0].refcount = 1;
- nv_wo32(chan->vm_pd, 0x0000, priv->pramin_pt->vinst | 0x63);
- nv_wo32(chan->vm_pd, 0x0004, 0);
+ nv50_instmem_map(dev_priv->bar3_vm->pgt[0].obj);
- /* DMA object for PRAMIN BAR */
- ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->pramin_bar);
+ ret = nv50_channel_new(dev, 128 * 1024, dev_priv->bar3_vm, &chan);
if (ret)
- return ret;
- nv_wo32(priv->pramin_bar, 0x00, 0x7fc00000);
- nv_wo32(priv->pramin_bar, 0x04, dev_priv->ramin_size - 1);
- nv_wo32(priv->pramin_bar, 0x08, 0x00000000);
- nv_wo32(priv->pramin_bar, 0x0c, 0x00000000);
- nv_wo32(priv->pramin_bar, 0x10, 0x00000000);
- nv_wo32(priv->pramin_bar, 0x14, 0x00000000);
-
- /* map channel into PRAMIN, gpuobj didn't do it for us */
- ret = nv50_instmem_bind(dev, chan->ramin);
+ goto error;
+ dev_priv->channels.ptr[0] = dev_priv->channels.ptr[127] = chan;
+
+ ret = nv50_gpuobj_dma_new(chan, 0x0000, BAR3_VM_BASE, BAR3_VM_SIZE,
+ NV_MEM_TARGET_VM, NV_MEM_ACCESS_VM,
+ NV_MEM_TYPE_VM, NV_MEM_COMP_VM,
+ &priv->bar3_dmaobj);
if (ret)
- return ret;
+ goto error;
- /* poke regs... */
nv_wr32(dev, 0x001704, 0x00000000 | (chan->ramin->vinst >> 12));
nv_wr32(dev, 0x001704, 0x40000000 | (chan->ramin->vinst >> 12));
- nv_wr32(dev, 0x00170c, 0x80000000 | (priv->pramin_bar->cinst >> 4));
+ nv_wr32(dev, 0x00170c, 0x80000000 | (priv->bar3_dmaobj->cinst >> 4));
tmp = nv_ri32(dev, 0);
nv_wi32(dev, 0, ~tmp);
if (nv_ri32(dev, 0) != ~tmp) {
NV_ERROR(dev, "PRAMIN readback failed\n");
- return -EIO;
+ ret = -EIO;
+ goto error;
}
nv_wi32(dev, 0, tmp);
dev_priv->ramin_available = true;
- /* Determine VM layout */
- dev_priv->vm_gart_base = roundup(NV50_VM_BLOCK, NV50_VM_BLOCK);
- dev_priv->vm_gart_size = NV50_VM_BLOCK;
-
- dev_priv->vm_vram_base = dev_priv->vm_gart_base + dev_priv->vm_gart_size;
- dev_priv->vm_vram_size = dev_priv->vram_size;
- if (dev_priv->vm_vram_size > NV50_VM_MAX_VRAM)
- dev_priv->vm_vram_size = NV50_VM_MAX_VRAM;
- dev_priv->vm_vram_size = roundup(dev_priv->vm_vram_size, NV50_VM_BLOCK);
- dev_priv->vm_vram_pt_nr = dev_priv->vm_vram_size / NV50_VM_BLOCK;
-
- dev_priv->vm_end = dev_priv->vm_vram_base + dev_priv->vm_vram_size;
-
- NV_DEBUG(dev, "NV50VM: GART 0x%016llx-0x%016llx\n",
- dev_priv->vm_gart_base,
- dev_priv->vm_gart_base + dev_priv->vm_gart_size - 1);
- NV_DEBUG(dev, "NV50VM: VRAM 0x%016llx-0x%016llx\n",
- dev_priv->vm_vram_base,
- dev_priv->vm_vram_base + dev_priv->vm_vram_size - 1);
-
- /* VRAM page table(s), mapped into VM at +1GiB */
- for (i = 0; i < dev_priv->vm_vram_pt_nr; i++) {
- ret = nouveau_gpuobj_new(dev, NULL, NV50_VM_BLOCK / 0x10000 * 8,
- 0, NVOBJ_FLAG_ZERO_ALLOC,
- &chan->vm_vram_pt[i]);
- if (ret) {
- NV_ERROR(dev, "Error creating VRAM PGT: %d\n", ret);
- dev_priv->vm_vram_pt_nr = i;
- return ret;
- }
- dev_priv->vm_vram_pt[i] = chan->vm_vram_pt[i];
-
- nv_wo32(chan->vm_pd, 0x10 + (i*8),
- chan->vm_vram_pt[i]->vinst | 0x61);
- nv_wo32(chan->vm_pd, 0x14 + (i*8), 0);
- }
+ /* BAR1 */
+ ret = nouveau_vm_new(dev, BAR1_VM_BASE, BAR1_VM_SIZE, BAR1_VM_BASE,
+ 29, 12, 16, &vm);
+ if (ret)
+ goto error;
- /* DMA object for FB BAR */
- ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->fb_bar);
+ ret = nouveau_vm_ref(vm, &dev_priv->bar1_vm, chan->vm_pd);
if (ret)
- return ret;
- nv_wo32(priv->fb_bar, 0x00, 0x7fc00000);
- nv_wo32(priv->fb_bar, 0x04, 0x40000000 +
- pci_resource_len(dev->pdev, 1) - 1);
- nv_wo32(priv->fb_bar, 0x08, 0x40000000);
- nv_wo32(priv->fb_bar, 0x0c, 0x00000000);
- nv_wo32(priv->fb_bar, 0x10, 0x00000000);
- nv_wo32(priv->fb_bar, 0x14, 0x00000000);
+ goto error;
+ nouveau_vm_ref(NULL, &vm, NULL);
- dev_priv->engine.instmem.flush(dev);
+ ret = nv50_gpuobj_dma_new(chan, 0x0000, BAR1_VM_BASE, BAR1_VM_SIZE,
+ NV_MEM_TARGET_VM, NV_MEM_ACCESS_VM,
+ NV_MEM_TYPE_VM, NV_MEM_COMP_VM,
+ &priv->bar1_dmaobj);
+ if (ret)
+ goto error;
- nv_wr32(dev, 0x001708, 0x80000000 | (priv->fb_bar->cinst >> 4));
+ nv_wr32(dev, 0x001708, 0x80000000 | (priv->bar1_dmaobj->cinst >> 4));
for (i = 0; i < 8; i++)
nv_wr32(dev, 0x1900 + (i*4), 0);
+ /* Create shared channel VM, space is reserved for GART mappings at
+ * the beginning of this address space, it's managed separately
+ * because TTM makes life painful
+ */
+ dev_priv->vm_gart_base = 0x0020000000ULL;
+ dev_priv->vm_gart_size = 512 * 1024 * 1024;
+ nongart_o = dev_priv->vm_gart_base + dev_priv->vm_gart_size;
+
+ ret = nouveau_vm_new(dev, 0, (1ULL << 40), nongart_o,
+ 29, 12, 16, &dev_priv->chan_vm);
+ if (ret)
+ return ret;
+
return 0;
+
+error:
+ nv50_instmem_takedown(dev);
+ return ret;
}
void
{
struct drm_nouveau_private *dev_priv = dev->dev_private;
struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
- struct nouveau_channel *chan = dev_priv->fifos[0];
+ struct nouveau_channel *chan = dev_priv->channels.ptr[0];
int i;
NV_DEBUG(dev, "\n");
dev_priv->ramin_available = false;
- /* Restore state from before init */
+ nouveau_vm_ref(NULL, &dev_priv->chan_vm, NULL);
+
for (i = 0x1700; i <= 0x1710; i += 4)
nv_wr32(dev, i, priv->save1700[(i - 0x1700) / 4]);
- nouveau_gpuobj_ref(NULL, &priv->fb_bar);
- nouveau_gpuobj_ref(NULL, &priv->pramin_bar);
- nouveau_gpuobj_ref(NULL, &priv->pramin_pt);
+ nouveau_gpuobj_ref(NULL, &priv->bar3_dmaobj);
+ nouveau_gpuobj_ref(NULL, &priv->bar1_dmaobj);
- /* Destroy dummy channel */
- if (chan) {
- for (i = 0; i < dev_priv->vm_vram_pt_nr; i++)
- nouveau_gpuobj_ref(NULL, &chan->vm_vram_pt[i]);
- dev_priv->vm_vram_pt_nr = 0;
+ nouveau_vm_ref(NULL, &dev_priv->bar1_vm, chan->vm_pd);
+ dev_priv->channels.ptr[127] = 0;
+ nv50_channel_del(&dev_priv->channels.ptr[0]);
- nv50_channel_del(&dev_priv->fifos[0]);
- dev_priv->fifos[127] = NULL;
- }
+ nouveau_gpuobj_ref(NULL, &dev_priv->bar3_vm->pgt[0].obj);
+ nouveau_vm_ref(NULL, &dev_priv->bar3_vm, NULL);
+
+ if (dev_priv->ramin_heap.free_stack.next)
+ drm_mm_takedown(&dev_priv->ramin_heap);
dev_priv->engine.instmem.priv = NULL;
kfree(priv);
nv50_instmem_suspend(struct drm_device *dev)
{
struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nouveau_channel *chan = dev_priv->fifos[0];
- struct nouveau_gpuobj *ramin = chan->ramin;
- int i;
-
- ramin->im_backing_suspend = vmalloc(ramin->size);
- if (!ramin->im_backing_suspend)
- return -ENOMEM;
- for (i = 0; i < ramin->size; i += 4)
- ramin->im_backing_suspend[i/4] = nv_ri32(dev, i);
+ dev_priv->ramin_available = false;
return 0;
}
{
struct drm_nouveau_private *dev_priv = dev->dev_private;
struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
- struct nouveau_channel *chan = dev_priv->fifos[0];
- struct nouveau_gpuobj *ramin = chan->ramin;
+ struct nouveau_channel *chan = dev_priv->channels.ptr[0];
int i;
- dev_priv->ramin_available = false;
- dev_priv->ramin_base = ~0;
- for (i = 0; i < ramin->size; i += 4)
- nv_wo32(ramin, i, ramin->im_backing_suspend[i/4]);
- dev_priv->ramin_available = true;
- vfree(ramin->im_backing_suspend);
- ramin->im_backing_suspend = NULL;
-
/* Poke the relevant regs, and pray it works :) */
nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12));
nv_wr32(dev, NV50_PUNK_UNK1710, 0);
nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12) |
NV50_PUNK_BAR_CFG_BASE_VALID);
- nv_wr32(dev, NV50_PUNK_BAR1_CTXDMA, (priv->fb_bar->cinst >> 4) |
+ nv_wr32(dev, NV50_PUNK_BAR1_CTXDMA, (priv->bar1_dmaobj->cinst >> 4) |
NV50_PUNK_BAR1_CTXDMA_VALID);
- nv_wr32(dev, NV50_PUNK_BAR3_CTXDMA, (priv->pramin_bar->cinst >> 4) |
+ nv_wr32(dev, NV50_PUNK_BAR3_CTXDMA, (priv->bar3_dmaobj->cinst >> 4) |
NV50_PUNK_BAR3_CTXDMA_VALID);
for (i = 0; i < 8; i++)
nv_wr32(dev, 0x1900 + (i*4), 0);
+
+ dev_priv->ramin_available = true;
}
+struct nv50_gpuobj_node {
+ struct nouveau_vram *vram;
+ u32 align;
+};
+
+
int
-nv50_instmem_populate(struct drm_device *dev, struct nouveau_gpuobj *gpuobj,
- uint32_t *sz)
+nv50_instmem_get(struct nouveau_gpuobj *gpuobj, u32 size, u32 align)
{
+ struct drm_device *dev = gpuobj->dev;
+ struct nv50_gpuobj_node *node = NULL;
int ret;
- if (gpuobj->im_backing)
- return -EINVAL;
-
- *sz = ALIGN(*sz, 4096);
- if (*sz == 0)
- return -EINVAL;
+ node = kzalloc(sizeof(*node), GFP_KERNEL);
+ if (!node)
+ return -ENOMEM;
+ node->align = align;
- ret = nouveau_bo_new(dev, NULL, *sz, 0, TTM_PL_FLAG_VRAM, 0, 0x0000,
- true, false, &gpuobj->im_backing);
- if (ret) {
- NV_ERROR(dev, "error getting PRAMIN backing pages: %d\n", ret);
- return ret;
- }
+ size = (size + 4095) & ~4095;
+ align = max(align, (u32)4096);
- ret = nouveau_bo_pin(gpuobj->im_backing, TTM_PL_FLAG_VRAM);
+ ret = nv50_vram_new(dev, size, align, 0, 0, &node->vram);
if (ret) {
- NV_ERROR(dev, "error pinning PRAMIN backing VRAM: %d\n", ret);
- nouveau_bo_ref(NULL, &gpuobj->im_backing);
+ kfree(node);
return ret;
}
- gpuobj->vinst = gpuobj->im_backing->bo.mem.start << PAGE_SHIFT;
+ gpuobj->vinst = node->vram->offset;
+ gpuobj->size = size;
+ gpuobj->node = node;
return 0;
}
void
-nv50_instmem_clear(struct drm_device *dev, struct nouveau_gpuobj *gpuobj)
+nv50_instmem_put(struct nouveau_gpuobj *gpuobj)
{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
+ struct drm_device *dev = gpuobj->dev;
+ struct nv50_gpuobj_node *node;
- if (gpuobj && gpuobj->im_backing) {
- if (gpuobj->im_bound)
- dev_priv->engine.instmem.unbind(dev, gpuobj);
- nouveau_bo_unpin(gpuobj->im_backing);
- nouveau_bo_ref(NULL, &gpuobj->im_backing);
- gpuobj->im_backing = NULL;
- }
+ node = gpuobj->node;
+ gpuobj->node = NULL;
+
+ nv50_vram_del(dev, &node->vram);
+ kfree(node);
}
int
-nv50_instmem_bind(struct drm_device *dev, struct nouveau_gpuobj *gpuobj)
+nv50_instmem_map(struct nouveau_gpuobj *gpuobj)
{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
- struct nouveau_gpuobj *pramin_pt = priv->pramin_pt;
- uint32_t pte, pte_end;
- uint64_t vram;
-
- if (!gpuobj->im_backing || !gpuobj->im_pramin || gpuobj->im_bound)
- return -EINVAL;
-
- NV_DEBUG(dev, "st=0x%lx sz=0x%lx\n",
- gpuobj->im_pramin->start, gpuobj->im_pramin->size);
-
- pte = (gpuobj->im_pramin->start >> 12) << 1;
- pte_end = ((gpuobj->im_pramin->size >> 12) << 1) + pte;
- vram = gpuobj->vinst;
-
- NV_DEBUG(dev, "pramin=0x%lx, pte=%d, pte_end=%d\n",
- gpuobj->im_pramin->start, pte, pte_end);
- NV_DEBUG(dev, "first vram page: 0x%010llx\n", gpuobj->vinst);
-
- vram |= 1;
- if (dev_priv->vram_sys_base) {
- vram += dev_priv->vram_sys_base;
- vram |= 0x30;
- }
-
- while (pte < pte_end) {
- nv_wo32(pramin_pt, (pte * 4) + 0, lower_32_bits(vram));
- nv_wo32(pramin_pt, (pte * 4) + 4, upper_32_bits(vram));
- vram += 0x1000;
- pte += 2;
- }
- dev_priv->engine.instmem.flush(dev);
+ struct drm_nouveau_private *dev_priv = gpuobj->dev->dev_private;
+ struct nv50_gpuobj_node *node = gpuobj->node;
+ int ret;
- nv50_vm_flush(dev, 4);
- nv50_vm_flush(dev, 6);
+ ret = nouveau_vm_get(dev_priv->bar3_vm, gpuobj->size, 12,
+ NV_MEM_ACCESS_RW, &node->vram->bar_vma);
+ if (ret)
+ return ret;
- gpuobj->im_bound = 1;
+ nouveau_vm_map(&node->vram->bar_vma, node->vram);
+ gpuobj->pinst = node->vram->bar_vma.offset;
return 0;
}
-int
-nv50_instmem_unbind(struct drm_device *dev, struct nouveau_gpuobj *gpuobj)
+void
+nv50_instmem_unmap(struct nouveau_gpuobj *gpuobj)
{
- struct drm_nouveau_private *dev_priv = dev->dev_private;
- struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
- uint32_t pte, pte_end;
+ struct nv50_gpuobj_node *node = gpuobj->node;
- if (gpuobj->im_bound == 0)
- return -EINVAL;
-
- /* can happen during late takedown */
- if (unlikely(!dev_priv->ramin_available))
- return 0;
-
- pte = (gpuobj->im_pramin->start >> 12) << 1;
- pte_end = ((gpuobj->im_pramin->size >> 12) << 1) + pte;
-
- while (pte < pte_end) {
- nv_wo32(priv->pramin_pt, (pte * 4) + 0, 0x00000000);
- nv_wo32(priv->pramin_pt, (pte * 4) + 4, 0x00000000);
- pte += 2;
+ if (node->vram->bar_vma.node) {
+ nouveau_vm_unmap(&node->vram->bar_vma);
+ nouveau_vm_put(&node->vram->bar_vma);
}
- dev_priv->engine.instmem.flush(dev);
-
- gpuobj->im_bound = 0;
- return 0;
}
void
NV_ERROR(dev, "PRAMIN flush timeout\n");
}
-void
-nv50_vm_flush(struct drm_device *dev, int engine)
-{
- nv_wr32(dev, 0x100c80, (engine << 16) | 1);
- if (!nv_wait(dev, 0x100c80, 0x00000001, 0x00000000))
- NV_ERROR(dev, "vm flush timeout: engine %d\n", engine);
-}
-
projects / cascardo / linux.git / blobdiff