2 #include "nouveau_drv.h"
3 #include <linux/pagemap.h>
4 #include <linux/slab.h>
6 #define NV_CTXDMA_PAGE_SHIFT 12
7 #define NV_CTXDMA_PAGE_SIZE (1 << NV_CTXDMA_PAGE_SHIFT)
8 #define NV_CTXDMA_PAGE_MASK (NV_CTXDMA_PAGE_SIZE - 1)
10 struct nouveau_sgdma_be {
11 struct ttm_backend backend;
12 struct drm_device *dev;
23 nouveau_sgdma_populate(struct ttm_backend *be, unsigned long num_pages,
24 struct page **pages, struct page *dummy_read_page,
25 dma_addr_t *dma_addrs)
27 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
28 struct drm_device *dev = nvbe->dev;
30 NV_DEBUG(nvbe->dev, "num_pages = %ld\n", num_pages);
35 nvbe->pages = kmalloc(sizeof(dma_addr_t) * num_pages, GFP_KERNEL);
39 nvbe->ttm_alloced = kmalloc(sizeof(bool) * num_pages, GFP_KERNEL);
40 if (!nvbe->ttm_alloced) {
48 /* this code path isn't called and is incorrect anyways */
49 if (0) { /*dma_addrs[nvbe->nr_pages] != DMA_ERROR_CODE)*/
50 nvbe->pages[nvbe->nr_pages] =
51 dma_addrs[nvbe->nr_pages];
52 nvbe->ttm_alloced[nvbe->nr_pages] = true;
54 nvbe->pages[nvbe->nr_pages] =
55 pci_map_page(dev->pdev, pages[nvbe->nr_pages], 0,
56 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
57 if (pci_dma_mapping_error(dev->pdev,
58 nvbe->pages[nvbe->nr_pages])) {
62 nvbe->ttm_alloced[nvbe->nr_pages] = false;
72 nouveau_sgdma_clear(struct ttm_backend *be)
74 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
75 struct drm_device *dev;
77 if (nvbe && nvbe->pages) {
84 while (nvbe->nr_pages--) {
85 if (!nvbe->ttm_alloced[nvbe->nr_pages])
86 pci_unmap_page(dev->pdev, nvbe->pages[nvbe->nr_pages],
87 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
90 kfree(nvbe->ttm_alloced);
92 nvbe->ttm_alloced = NULL;
98 nouveau_sgdma_destroy(struct ttm_backend *be)
100 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
103 NV_DEBUG(nvbe->dev, "\n");
114 nv04_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
116 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
117 struct drm_device *dev = nvbe->dev;
118 struct drm_nouveau_private *dev_priv = dev->dev_private;
119 struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
122 NV_DEBUG(dev, "pg=0x%lx\n", mem->start);
124 nvbe->offset = mem->start << PAGE_SHIFT;
125 pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
126 for (i = 0; i < nvbe->nr_pages; i++) {
127 dma_addr_t dma_offset = nvbe->pages[i];
128 uint32_t offset_l = lower_32_bits(dma_offset);
130 for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++, pte++) {
131 nv_wo32(gpuobj, (pte * 4) + 0, offset_l | 3);
132 offset_l += NV_CTXDMA_PAGE_SIZE;
141 nv04_sgdma_unbind(struct ttm_backend *be)
143 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
144 struct drm_device *dev = nvbe->dev;
145 struct drm_nouveau_private *dev_priv = dev->dev_private;
146 struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
154 pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
155 for (i = 0; i < nvbe->nr_pages; i++) {
156 for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++, pte++)
157 nv_wo32(gpuobj, (pte * 4) + 0, 0x00000000);
164 static struct ttm_backend_func nv04_sgdma_backend = {
165 .populate = nouveau_sgdma_populate,
166 .clear = nouveau_sgdma_clear,
167 .bind = nv04_sgdma_bind,
168 .unbind = nv04_sgdma_unbind,
169 .destroy = nouveau_sgdma_destroy
173 nv41_sgdma_flush(struct nouveau_sgdma_be *nvbe)
175 struct drm_device *dev = nvbe->dev;
177 nv_wr32(dev, 0x100810, 0x00000022);
178 if (!nv_wait(dev, 0x100810, 0x00000100, 0x00000100))
179 NV_ERROR(dev, "vm flush timeout: 0x%08x\n",
180 nv_rd32(dev, 0x100810));
181 nv_wr32(dev, 0x100810, 0x00000000);
185 nv41_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
187 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
188 struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
189 struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
190 dma_addr_t *list = nvbe->pages;
191 u32 pte = mem->start << 2;
192 u32 cnt = nvbe->nr_pages;
194 nvbe->offset = mem->start << PAGE_SHIFT;
197 nv_wo32(pgt, pte, (*list++ >> 7) | 1);
201 nv41_sgdma_flush(nvbe);
207 nv41_sgdma_unbind(struct ttm_backend *be)
209 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
210 struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
211 struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
212 u32 pte = (nvbe->offset >> 12) << 2;
213 u32 cnt = nvbe->nr_pages;
216 nv_wo32(pgt, pte, 0x00000000);
220 nv41_sgdma_flush(nvbe);
225 static struct ttm_backend_func nv41_sgdma_backend = {
226 .populate = nouveau_sgdma_populate,
227 .clear = nouveau_sgdma_clear,
228 .bind = nv41_sgdma_bind,
229 .unbind = nv41_sgdma_unbind,
230 .destroy = nouveau_sgdma_destroy
234 nv44_sgdma_flush(struct nouveau_sgdma_be *nvbe)
236 struct drm_device *dev = nvbe->dev;
238 nv_wr32(dev, 0x100814, (nvbe->nr_pages - 1) << 12);
239 nv_wr32(dev, 0x100808, nvbe->offset | 0x20);
240 if (!nv_wait(dev, 0x100808, 0x00000001, 0x00000001))
241 NV_ERROR(dev, "gart flush timeout: 0x%08x\n",
242 nv_rd32(dev, 0x100808));
243 nv_wr32(dev, 0x100808, 0x00000000);
247 nv44_sgdma_fill(struct nouveau_gpuobj *pgt, dma_addr_t *list, u32 base, u32 cnt)
249 struct drm_nouveau_private *dev_priv = pgt->dev->dev_private;
250 dma_addr_t dummy = dev_priv->gart_info.dummy.addr;
256 tmp[0] = nv_ro32(pgt, base + 0x0);
257 tmp[1] = nv_ro32(pgt, base + 0x4);
258 tmp[2] = nv_ro32(pgt, base + 0x8);
259 tmp[3] = nv_ro32(pgt, base + 0xc);
261 u32 addr = list ? (*list++ >> 12) : (dummy >> 12);
262 switch (pte++ & 0x3) {
264 tmp[0] &= ~0x07ffffff;
268 tmp[0] &= ~0xf8000000;
269 tmp[0] |= addr << 27;
270 tmp[1] &= ~0x003fffff;
274 tmp[1] &= ~0xffc00000;
275 tmp[1] |= addr << 22;
276 tmp[2] &= ~0x0001ffff;
277 tmp[2] |= addr >> 10;
280 tmp[2] &= ~0xfffe0000;
281 tmp[2] |= addr << 17;
282 tmp[3] &= ~0x00000fff;
283 tmp[3] |= addr >> 15;
288 tmp[3] |= 0x40000000;
290 nv_wo32(pgt, base + 0x0, tmp[0]);
291 nv_wo32(pgt, base + 0x4, tmp[1]);
292 nv_wo32(pgt, base + 0x8, tmp[2]);
293 nv_wo32(pgt, base + 0xc, tmp[3]);
297 nv44_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
299 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
300 struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
301 struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
302 dma_addr_t *list = nvbe->pages;
303 u32 pte = mem->start << 2, tmp[4];
304 u32 cnt = nvbe->nr_pages;
307 nvbe->offset = mem->start << PAGE_SHIFT;
309 if (pte & 0x0000000c) {
310 u32 max = 4 - ((pte >> 2) & 0x3);
311 u32 part = (cnt > max) ? max : cnt;
312 nv44_sgdma_fill(pgt, list, pte, part);
319 for (i = 0; i < 4; i++)
320 tmp[i] = *list++ >> 12;
321 nv_wo32(pgt, pte + 0x0, tmp[0] >> 0 | tmp[1] << 27);
322 nv_wo32(pgt, pte + 0x4, tmp[1] >> 5 | tmp[2] << 22);
323 nv_wo32(pgt, pte + 0x8, tmp[2] >> 10 | tmp[3] << 17);
324 nv_wo32(pgt, pte + 0xc, tmp[3] >> 15 | 0x40000000);
330 nv44_sgdma_fill(pgt, list, pte, cnt);
332 nv44_sgdma_flush(nvbe);
338 nv44_sgdma_unbind(struct ttm_backend *be)
340 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
341 struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
342 struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
343 u32 pte = (nvbe->offset >> 12) << 2;
344 u32 cnt = nvbe->nr_pages;
346 if (pte & 0x0000000c) {
347 u32 max = 4 - ((pte >> 2) & 0x3);
348 u32 part = (cnt > max) ? max : cnt;
349 nv44_sgdma_fill(pgt, NULL, pte, part);
355 nv_wo32(pgt, pte + 0x0, 0x00000000);
356 nv_wo32(pgt, pte + 0x4, 0x00000000);
357 nv_wo32(pgt, pte + 0x8, 0x00000000);
358 nv_wo32(pgt, pte + 0xc, 0x00000000);
364 nv44_sgdma_fill(pgt, NULL, pte, cnt);
366 nv44_sgdma_flush(nvbe);
371 static struct ttm_backend_func nv44_sgdma_backend = {
372 .populate = nouveau_sgdma_populate,
373 .clear = nouveau_sgdma_clear,
374 .bind = nv44_sgdma_bind,
375 .unbind = nv44_sgdma_unbind,
376 .destroy = nouveau_sgdma_destroy
380 nv50_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
382 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
383 struct nouveau_mem *node = mem->mm_node;
384 /* noop: bound in move_notify() */
385 node->pages = nvbe->pages;
386 nvbe->pages = (dma_addr_t *)node;
392 nv50_sgdma_unbind(struct ttm_backend *be)
394 struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
395 struct nouveau_mem *node = (struct nouveau_mem *)nvbe->pages;
396 /* noop: unbound in move_notify() */
397 nvbe->pages = node->pages;
403 static struct ttm_backend_func nv50_sgdma_backend = {
404 .populate = nouveau_sgdma_populate,
405 .clear = nouveau_sgdma_clear,
406 .bind = nv50_sgdma_bind,
407 .unbind = nv50_sgdma_unbind,
408 .destroy = nouveau_sgdma_destroy
412 nouveau_sgdma_init_ttm(struct drm_device *dev)
414 struct drm_nouveau_private *dev_priv = dev->dev_private;
415 struct nouveau_sgdma_be *nvbe;
417 nvbe = kzalloc(sizeof(*nvbe), GFP_KERNEL);
423 nvbe->backend.func = dev_priv->gart_info.func;
424 return &nvbe->backend;
428 nouveau_sgdma_init(struct drm_device *dev)
430 struct drm_nouveau_private *dev_priv = dev->dev_private;
431 struct nouveau_gpuobj *gpuobj = NULL;
432 u32 aper_size, align;
435 if (dev_priv->card_type >= NV_40 && pci_is_pcie(dev->pdev))
436 aper_size = 512 * 1024 * 1024;
438 aper_size = 64 * 1024 * 1024;
440 /* Dear NVIDIA, NV44+ would like proper present bits in PTEs for
441 * christmas. The cards before it have them, the cards after
442 * it have them, why is NV44 so unloved?
444 dev_priv->gart_info.dummy.page = alloc_page(GFP_DMA32 | GFP_KERNEL);
445 if (!dev_priv->gart_info.dummy.page)
448 dev_priv->gart_info.dummy.addr =
449 pci_map_page(dev->pdev, dev_priv->gart_info.dummy.page,
450 0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
451 if (pci_dma_mapping_error(dev->pdev, dev_priv->gart_info.dummy.addr)) {
452 NV_ERROR(dev, "error mapping dummy page\n");
453 __free_page(dev_priv->gart_info.dummy.page);
454 dev_priv->gart_info.dummy.page = NULL;
458 if (dev_priv->card_type >= NV_50) {
459 dev_priv->gart_info.aper_base = 0;
460 dev_priv->gart_info.aper_size = aper_size;
461 dev_priv->gart_info.type = NOUVEAU_GART_HW;
462 dev_priv->gart_info.func = &nv50_sgdma_backend;
464 if (0 && pci_is_pcie(dev->pdev) &&
465 dev_priv->chipset > 0x40 && dev_priv->chipset != 0x45) {
466 if (nv44_graph_class(dev)) {
467 dev_priv->gart_info.func = &nv44_sgdma_backend;
470 dev_priv->gart_info.func = &nv41_sgdma_backend;
474 ret = nouveau_gpuobj_new(dev, NULL, aper_size / 1024, align,
475 NVOBJ_FLAG_ZERO_ALLOC |
476 NVOBJ_FLAG_ZERO_FREE, &gpuobj);
478 NV_ERROR(dev, "Error creating sgdma object: %d\n", ret);
482 dev_priv->gart_info.sg_ctxdma = gpuobj;
483 dev_priv->gart_info.aper_base = 0;
484 dev_priv->gart_info.aper_size = aper_size;
485 dev_priv->gart_info.type = NOUVEAU_GART_HW;
487 ret = nouveau_gpuobj_new(dev, NULL, (aper_size / 1024) + 8, 16,
488 NVOBJ_FLAG_ZERO_ALLOC |
489 NVOBJ_FLAG_ZERO_FREE, &gpuobj);
491 NV_ERROR(dev, "Error creating sgdma object: %d\n", ret);
495 nv_wo32(gpuobj, 0, NV_CLASS_DMA_IN_MEMORY |
496 (1 << 12) /* PT present */ |
497 (0 << 13) /* PT *not* linear */ |
499 (2 << 16) /* PCI */);
500 nv_wo32(gpuobj, 4, aper_size - 1);
502 dev_priv->gart_info.sg_ctxdma = gpuobj;
503 dev_priv->gart_info.aper_base = 0;
504 dev_priv->gart_info.aper_size = aper_size;
505 dev_priv->gart_info.type = NOUVEAU_GART_PDMA;
506 dev_priv->gart_info.func = &nv04_sgdma_backend;
513 nouveau_sgdma_takedown(struct drm_device *dev)
515 struct drm_nouveau_private *dev_priv = dev->dev_private;
517 nouveau_gpuobj_ref(NULL, &dev_priv->gart_info.sg_ctxdma);
519 if (dev_priv->gart_info.dummy.page) {
520 pci_unmap_page(dev->pdev, dev_priv->gart_info.dummy.addr,
521 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
522 __free_page(dev_priv->gart_info.dummy.page);
523 dev_priv->gart_info.dummy.page = NULL;
528 nouveau_sgdma_get_physical(struct drm_device *dev, uint32_t offset)
530 struct drm_nouveau_private *dev_priv = dev->dev_private;
531 struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
532 int pte = (offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
534 BUG_ON(dev_priv->card_type >= NV_50);
536 return (nv_ro32(gpuobj, 4 * pte) & ~NV_CTXDMA_PAGE_MASK) |
537 (offset & NV_CTXDMA_PAGE_MASK);