drm/nv50: allow gpuobjs that aren't mapped into aperture
[linux-2.6.git] / drivers / gpu / drm / nouveau / nv50_instmem.c
blob5c617f807b237a64976f07a849c35400e7cb29b5
1 /*
2 * Copyright (C) 2007 Ben Skeggs.
4 * All Rights Reserved.
6 * Permission is hereby granted, free of charge, to any person obtaining
7 * a copy of this software and associated documentation files (the
8 * "Software"), to deal in the Software without restriction, including
9 * without limitation the rights to use, copy, modify, merge, publish,
10 * distribute, sublicense, and/or sell copies of the Software, and to
11 * permit persons to whom the Software is furnished to do so, subject to
12 * the following conditions:
14 * The above copyright notice and this permission notice (including the
15 * next paragraph) shall be included in all copies or substantial
16 * portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 * IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 * LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 * OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 * WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
28 #include "drmP.h"
29 #include "drm.h"
30 #include "nouveau_drv.h"
32 struct nv50_instmem_priv {
33 uint32_t save1700[5]; /* 0x1700->0x1710 */
35 struct nouveau_gpuobj *pramin_pt;
36 struct nouveau_gpuobj *pramin_bar;
37 struct nouveau_gpuobj *fb_bar;
40 #define NV50_INSTMEM_PAGE_SHIFT 12
41 #define NV50_INSTMEM_PAGE_SIZE (1 << NV50_INSTMEM_PAGE_SHIFT)
42 #define NV50_INSTMEM_PT_SIZE(a) (((a) >> 12) << 3)
44 /*NOTE: - Assumes 0x1700 already covers the correct MiB of PRAMIN
46 #define BAR0_WI32(g, o, v) do { \
47 u32 offset = (g)->vinst + (o); \
48 nv_wr32(dev, NV_RAMIN + (offset & 0xfffff), (v)); \
49 } while (0)
51 int
52 nv50_instmem_init(struct drm_device *dev)
54 struct drm_nouveau_private *dev_priv = dev->dev_private;
55 struct nouveau_channel *chan;
56 uint32_t c_offset, c_size, c_ramfc, c_vmpd, c_base, pt_size;
57 uint32_t save_nv001700;
58 uint64_t v;
59 struct nv50_instmem_priv *priv;
60 int ret, i;
62 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
63 if (!priv)
64 return -ENOMEM;
65 dev_priv->engine.instmem.priv = priv;
67 /* Save state, will restore at takedown. */
68 for (i = 0x1700; i <= 0x1710; i += 4)
69 priv->save1700[(i-0x1700)/4] = nv_rd32(dev, i);
71 /* Reserve the last MiB of VRAM, we should probably try to avoid
72 * setting up the below tables over the top of the VBIOS image at
73 * some point.
75 dev_priv->ramin_rsvd_vram = 1 << 20;
76 c_offset = dev_priv->vram_size - dev_priv->ramin_rsvd_vram;
77 c_size = 128 << 10;
78 c_vmpd = ((dev_priv->chipset & 0xf0) == 0x50) ? 0x1400 : 0x200;
79 c_ramfc = ((dev_priv->chipset & 0xf0) == 0x50) ? 0x0 : 0x20;
80 c_base = c_vmpd + 0x4000;
81 pt_size = NV50_INSTMEM_PT_SIZE(dev_priv->ramin_size);
83 NV_DEBUG(dev, " Rsvd VRAM base: 0x%08x\n", c_offset);
84 NV_DEBUG(dev, " VBIOS image: 0x%08x\n",
85 (nv_rd32(dev, 0x619f04) & ~0xff) << 8);
86 NV_DEBUG(dev, " Aperture size: %d MiB\n", dev_priv->ramin_size >> 20);
87 NV_DEBUG(dev, " PT size: %d KiB\n", pt_size >> 10);
89 /* Determine VM layout, we need to do this first to make sure
90 * we allocate enough memory for all the page tables.
92 dev_priv->vm_gart_base = roundup(NV50_VM_BLOCK, NV50_VM_BLOCK);
93 dev_priv->vm_gart_size = NV50_VM_BLOCK;
95 dev_priv->vm_vram_base = dev_priv->vm_gart_base + dev_priv->vm_gart_size;
96 dev_priv->vm_vram_size = dev_priv->vram_size;
97 if (dev_priv->vm_vram_size > NV50_VM_MAX_VRAM)
98 dev_priv->vm_vram_size = NV50_VM_MAX_VRAM;
99 dev_priv->vm_vram_size = roundup(dev_priv->vm_vram_size, NV50_VM_BLOCK);
100 dev_priv->vm_vram_pt_nr = dev_priv->vm_vram_size / NV50_VM_BLOCK;
102 dev_priv->vm_end = dev_priv->vm_vram_base + dev_priv->vm_vram_size;
104 NV_DEBUG(dev, "NV50VM: GART 0x%016llx-0x%016llx\n",
105 dev_priv->vm_gart_base,
106 dev_priv->vm_gart_base + dev_priv->vm_gart_size - 1);
107 NV_DEBUG(dev, "NV50VM: VRAM 0x%016llx-0x%016llx\n",
108 dev_priv->vm_vram_base,
109 dev_priv->vm_vram_base + dev_priv->vm_vram_size - 1);
111 c_size += dev_priv->vm_vram_pt_nr * (NV50_VM_BLOCK / 65536 * 8);
113 /* Map BAR0 PRAMIN aperture over the memory we want to use */
114 save_nv001700 = nv_rd32(dev, NV50_PUNK_BAR0_PRAMIN);
115 nv_wr32(dev, NV50_PUNK_BAR0_PRAMIN, (c_offset >> 16));
117 /* Create a fake channel, and use it as our "dummy" channels 0/127.
118 * The main reason for creating a channel is so we can use the gpuobj
119 * code. However, it's probably worth noting that NVIDIA also setup
120 * their channels 0/127 with the same values they configure here.
121 * So, there may be some other reason for doing this.
123 * Have to create the entire channel manually, as the real channel
124 * creation code assumes we have PRAMIN access, and we don't until
125 * we're done here.
127 chan = kzalloc(sizeof(*chan), GFP_KERNEL);
128 if (!chan)
129 return -ENOMEM;
130 chan->id = 0;
131 chan->dev = dev;
132 chan->file_priv = (struct drm_file *)-2;
133 dev_priv->fifos[0] = dev_priv->fifos[127] = chan;
135 INIT_LIST_HEAD(&chan->ramht_refs);
137 /* Channel's PRAMIN object + heap */
138 ret = nouveau_gpuobj_new_fake(dev, 0, c_offset, c_size, 0, &chan->ramin);
139 if (ret)
140 return ret;
142 if (drm_mm_init(&chan->ramin_heap, c_base, c_size - c_base))
143 return -ENOMEM;
145 /* RAMFC + zero channel's PRAMIN up to start of VM pagedir */
146 ret = nouveau_gpuobj_new_fake(dev, c_ramfc, c_offset + c_ramfc,
147 0x4000, 0, &chan->ramfc);
148 if (ret)
149 return ret;
151 for (i = 0; i < c_vmpd; i += 4)
152 BAR0_WI32(chan->ramin, i, 0);
154 /* VM page directory */
155 ret = nouveau_gpuobj_new_fake(dev, c_vmpd, c_offset + c_vmpd,
156 0x4000, 0, &chan->vm_pd);
157 if (ret)
158 return ret;
159 for (i = 0; i < 0x4000; i += 8) {
160 BAR0_WI32(chan->vm_pd, i + 0x00, 0x00000000);
161 BAR0_WI32(chan->vm_pd, i + 0x04, 0x00000000);
164 /* PRAMIN page table, cheat and map into VM at 0x0000000000.
165 * We map the entire fake channel into the start of the PRAMIN BAR
167 ret = nouveau_gpuobj_new(dev, chan, pt_size, 0x1000, 0,
168 &priv->pramin_pt);
169 if (ret)
170 return ret;
172 v = c_offset | 1;
173 if (dev_priv->vram_sys_base) {
174 v += dev_priv->vram_sys_base;
175 v |= 0x30;
178 i = 0;
179 while (v < dev_priv->vram_sys_base + c_offset + c_size) {
180 BAR0_WI32(priv->pramin_pt, i + 0, lower_32_bits(v));
181 BAR0_WI32(priv->pramin_pt, i + 4, upper_32_bits(v));
182 v += 0x1000;
183 i += 8;
186 while (i < pt_size) {
187 BAR0_WI32(priv->pramin_pt, i + 0, 0x00000000);
188 BAR0_WI32(priv->pramin_pt, i + 4, 0x00000000);
189 i += 8;
192 BAR0_WI32(chan->vm_pd, 0x00, priv->pramin_pt->vinst | 0x63);
193 BAR0_WI32(chan->vm_pd, 0x04, 0x00000000);
195 /* VRAM page table(s), mapped into VM at +1GiB */
196 for (i = 0; i < dev_priv->vm_vram_pt_nr; i++) {
197 ret = nouveau_gpuobj_new(dev, chan, NV50_VM_BLOCK / 0x10000 * 8,
198 0, 0, &chan->vm_vram_pt[i]);
199 if (ret) {
200 NV_ERROR(dev, "Error creating VRAM page tables: %d\n",
201 ret);
202 dev_priv->vm_vram_pt_nr = i;
203 return ret;
205 /*XXX: double-check this is ok */
206 dev_priv->vm_vram_pt[i] = chan->vm_vram_pt[i];
208 for (v = 0; v < dev_priv->vm_vram_pt[i]->size; v += 4)
209 BAR0_WI32(dev_priv->vm_vram_pt[i], v, 0);
211 BAR0_WI32(chan->vm_pd, 0x10 + (i*8),
212 chan->vm_vram_pt[i]->vinst | 0x61);
213 BAR0_WI32(chan->vm_pd, 0x14 + (i*8), 0);
216 /* DMA object for PRAMIN BAR */
217 ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->pramin_bar);
218 if (ret)
219 return ret;
220 BAR0_WI32(priv->pramin_bar, 0x00, 0x7fc00000);
221 BAR0_WI32(priv->pramin_bar, 0x04, dev_priv->ramin_size - 1);
222 BAR0_WI32(priv->pramin_bar, 0x08, 0x00000000);
223 BAR0_WI32(priv->pramin_bar, 0x0c, 0x00000000);
224 BAR0_WI32(priv->pramin_bar, 0x10, 0x00000000);
225 BAR0_WI32(priv->pramin_bar, 0x14, 0x00000000);
227 /* DMA object for FB BAR */
228 ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->fb_bar);
229 if (ret)
230 return ret;
231 BAR0_WI32(priv->fb_bar, 0x00, 0x7fc00000);
232 BAR0_WI32(priv->fb_bar, 0x04, 0x40000000 +
233 pci_resource_len(dev->pdev, 1) - 1);
234 BAR0_WI32(priv->fb_bar, 0x08, 0x40000000);
235 BAR0_WI32(priv->fb_bar, 0x0c, 0x00000000);
236 BAR0_WI32(priv->fb_bar, 0x10, 0x00000000);
237 BAR0_WI32(priv->fb_bar, 0x14, 0x00000000);
239 /* Poke the relevant regs, and pray it works :) */
240 nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12));
241 nv_wr32(dev, NV50_PUNK_UNK1710, 0);
242 nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12) |
243 NV50_PUNK_BAR_CFG_BASE_VALID);
244 nv_wr32(dev, NV50_PUNK_BAR1_CTXDMA, (priv->fb_bar->cinst >> 4) |
245 NV50_PUNK_BAR1_CTXDMA_VALID);
246 nv_wr32(dev, NV50_PUNK_BAR3_CTXDMA, (priv->pramin_bar->cinst >> 4) |
247 NV50_PUNK_BAR3_CTXDMA_VALID);
249 for (i = 0; i < 8; i++)
250 nv_wr32(dev, 0x1900 + (i*4), 0);
252 dev_priv->ramin_available = true;
254 /* Assume that praying isn't enough, check that we can re-read the
255 * entire fake channel back from the PRAMIN BAR */
256 for (i = 0; i < c_size; i += 4) {
257 if (nv_rd32(dev, NV_RAMIN + i) != nv_ri32(dev, i)) {
258 NV_ERROR(dev, "Error reading back PRAMIN at 0x%08x\n",
260 return -EINVAL;
264 nv_wr32(dev, NV50_PUNK_BAR0_PRAMIN, save_nv001700);
266 /* Global PRAMIN heap */
267 if (drm_mm_init(&dev_priv->ramin_heap, c_size, dev_priv->ramin_size - c_size)) {
268 NV_ERROR(dev, "Failed to init RAMIN heap\n");
271 /*XXX: incorrect, but needed to make hash func "work" */
272 dev_priv->ramht_offset = 0x10000;
273 dev_priv->ramht_bits = 9;
274 dev_priv->ramht_size = (1 << dev_priv->ramht_bits) * 8;
275 return 0;
278 void
279 nv50_instmem_takedown(struct drm_device *dev)
281 struct drm_nouveau_private *dev_priv = dev->dev_private;
282 struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
283 struct nouveau_channel *chan = dev_priv->fifos[0];
284 int i;
286 NV_DEBUG(dev, "\n");
288 if (!priv)
289 return;
291 /* Restore state from before init */
292 for (i = 0x1700; i <= 0x1710; i += 4)
293 nv_wr32(dev, i, priv->save1700[(i - 0x1700) / 4]);
295 nouveau_gpuobj_ref(NULL, &priv->fb_bar);
296 nouveau_gpuobj_ref(NULL, &priv->pramin_bar);
297 nouveau_gpuobj_ref(NULL, &priv->pramin_pt);
299 /* Destroy dummy channel */
300 if (chan) {
301 for (i = 0; i < dev_priv->vm_vram_pt_nr; i++)
302 nouveau_gpuobj_ref(NULL, &chan->vm_vram_pt[i]);
303 dev_priv->vm_vram_pt_nr = 0;
305 nouveau_gpuobj_ref(NULL, &chan->vm_pd);
306 nouveau_gpuobj_ref(NULL, &chan->ramfc);
307 nouveau_gpuobj_ref(NULL, &chan->ramin);
308 drm_mm_takedown(&chan->ramin_heap);
310 dev_priv->fifos[0] = dev_priv->fifos[127] = NULL;
311 kfree(chan);
314 dev_priv->engine.instmem.priv = NULL;
315 kfree(priv);
319 nv50_instmem_suspend(struct drm_device *dev)
321 struct drm_nouveau_private *dev_priv = dev->dev_private;
322 struct nouveau_channel *chan = dev_priv->fifos[0];
323 struct nouveau_gpuobj *ramin = chan->ramin;
324 int i;
326 ramin->im_backing_suspend = vmalloc(ramin->size);
327 if (!ramin->im_backing_suspend)
328 return -ENOMEM;
330 for (i = 0; i < ramin->size; i += 4)
331 ramin->im_backing_suspend[i/4] = nv_ri32(dev, i);
332 return 0;
335 void
336 nv50_instmem_resume(struct drm_device *dev)
338 struct drm_nouveau_private *dev_priv = dev->dev_private;
339 struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
340 struct nouveau_channel *chan = dev_priv->fifos[0];
341 struct nouveau_gpuobj *ramin = chan->ramin;
342 int i;
344 nv_wr32(dev, NV50_PUNK_BAR0_PRAMIN, (ramin->vinst >> 16));
345 for (i = 0; i < ramin->size; i += 4)
346 BAR0_WI32(ramin, i, ramin->im_backing_suspend[i/4]);
347 vfree(ramin->im_backing_suspend);
348 ramin->im_backing_suspend = NULL;
350 /* Poke the relevant regs, and pray it works :) */
351 nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12));
352 nv_wr32(dev, NV50_PUNK_UNK1710, 0);
353 nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12) |
354 NV50_PUNK_BAR_CFG_BASE_VALID);
355 nv_wr32(dev, NV50_PUNK_BAR1_CTXDMA, (priv->fb_bar->cinst >> 4) |
356 NV50_PUNK_BAR1_CTXDMA_VALID);
357 nv_wr32(dev, NV50_PUNK_BAR3_CTXDMA, (priv->pramin_bar->cinst >> 4) |
358 NV50_PUNK_BAR3_CTXDMA_VALID);
360 for (i = 0; i < 8; i++)
361 nv_wr32(dev, 0x1900 + (i*4), 0);
365 nv50_instmem_populate(struct drm_device *dev, struct nouveau_gpuobj *gpuobj,
366 uint32_t *sz)
368 int ret;
370 if (gpuobj->im_backing)
371 return -EINVAL;
373 *sz = ALIGN(*sz, NV50_INSTMEM_PAGE_SIZE);
374 if (*sz == 0)
375 return -EINVAL;
377 ret = nouveau_bo_new(dev, NULL, *sz, 0, TTM_PL_FLAG_VRAM, 0, 0x0000,
378 true, false, &gpuobj->im_backing);
379 if (ret) {
380 NV_ERROR(dev, "error getting PRAMIN backing pages: %d\n", ret);
381 return ret;
384 ret = nouveau_bo_pin(gpuobj->im_backing, TTM_PL_FLAG_VRAM);
385 if (ret) {
386 NV_ERROR(dev, "error pinning PRAMIN backing VRAM: %d\n", ret);
387 nouveau_bo_ref(NULL, &gpuobj->im_backing);
388 return ret;
391 gpuobj->vinst = gpuobj->im_backing->bo.mem.mm_node->start << PAGE_SHIFT;
392 return 0;
395 void
396 nv50_instmem_clear(struct drm_device *dev, struct nouveau_gpuobj *gpuobj)
398 struct drm_nouveau_private *dev_priv = dev->dev_private;
400 if (gpuobj && gpuobj->im_backing) {
401 if (gpuobj->im_bound)
402 dev_priv->engine.instmem.unbind(dev, gpuobj);
403 nouveau_bo_unpin(gpuobj->im_backing);
404 nouveau_bo_ref(NULL, &gpuobj->im_backing);
405 gpuobj->im_backing = NULL;
410 nv50_instmem_bind(struct drm_device *dev, struct nouveau_gpuobj *gpuobj)
412 struct drm_nouveau_private *dev_priv = dev->dev_private;
413 struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
414 struct nouveau_gpuobj *pramin_pt = priv->pramin_pt;
415 uint32_t pte, pte_end;
416 uint64_t vram;
418 if (!gpuobj->im_backing || !gpuobj->im_pramin || gpuobj->im_bound)
419 return -EINVAL;
421 NV_DEBUG(dev, "st=0x%lx sz=0x%lx\n",
422 gpuobj->im_pramin->start, gpuobj->im_pramin->size);
424 pte = (gpuobj->im_pramin->start >> 12) << 1;
425 pte_end = ((gpuobj->im_pramin->size >> 12) << 1) + pte;
426 vram = gpuobj->vinst;
428 NV_DEBUG(dev, "pramin=0x%lx, pte=%d, pte_end=%d\n",
429 gpuobj->im_pramin->start, pte, pte_end);
430 NV_DEBUG(dev, "first vram page: 0x%010llx\n", gpuobj->vinst);
432 vram |= 1;
433 if (dev_priv->vram_sys_base) {
434 vram += dev_priv->vram_sys_base;
435 vram |= 0x30;
438 while (pte < pte_end) {
439 nv_wo32(pramin_pt, (pte * 4) + 0, lower_32_bits(vram));
440 nv_wo32(pramin_pt, (pte * 4) + 4, upper_32_bits(vram));
441 vram += NV50_INSTMEM_PAGE_SIZE;
442 pte += 2;
444 dev_priv->engine.instmem.flush(dev);
446 nv50_vm_flush(dev, 4);
447 nv50_vm_flush(dev, 6);
449 gpuobj->im_bound = 1;
450 return 0;
454 nv50_instmem_unbind(struct drm_device *dev, struct nouveau_gpuobj *gpuobj)
456 struct drm_nouveau_private *dev_priv = dev->dev_private;
457 struct nv50_instmem_priv *priv = dev_priv->engine.instmem.priv;
458 uint32_t pte, pte_end;
460 if (gpuobj->im_bound == 0)
461 return -EINVAL;
463 pte = (gpuobj->im_pramin->start >> 12) << 1;
464 pte_end = ((gpuobj->im_pramin->size >> 12) << 1) + pte;
466 while (pte < pte_end) {
467 nv_wo32(priv->pramin_pt, (pte * 4) + 0, 0x00000000);
468 nv_wo32(priv->pramin_pt, (pte * 4) + 4, 0x00000000);
469 pte += 2;
471 dev_priv->engine.instmem.flush(dev);
473 gpuobj->im_bound = 0;
474 return 0;
477 void
478 nv50_instmem_flush(struct drm_device *dev)
480 nv_wr32(dev, 0x00330c, 0x00000001);
481 if (!nv_wait(0x00330c, 0x00000002, 0x00000000))
482 NV_ERROR(dev, "PRAMIN flush timeout\n");
485 void
486 nv84_instmem_flush(struct drm_device *dev)
488 nv_wr32(dev, 0x070000, 0x00000001);
489 if (!nv_wait(0x070000, 0x00000002, 0x00000000))
490 NV_ERROR(dev, "PRAMIN flush timeout\n");
493 void
494 nv50_vm_flush(struct drm_device *dev, int engine)
496 nv_wr32(dev, 0x100c80, (engine << 16) | 1);
497 if (!nv_wait(0x100c80, 0x00000001, 0x00000000))
498 NV_ERROR(dev, "vm flush timeout: engine %d\n", engine);