1 #include "pipe/p_defines.h"
2 #include "pipe/p_screen.h"
3 #include "pipe/p_state.h"
4
5 #include "util/u_memory.h"
6 #include "util/u_inlines.h"
7 #include "util/format/u_format.h"
8 #include "util/format/u_format_s3tc.h"
9 #include "util/u_string.h"
10
11 #include "os/os_mman.h"
12 #include "util/os_time.h"
13
14 #include <stdio.h>
15 #include <errno.h>
16 #include <stdlib.h>
17
18 #include <nouveau_drm.h>
19 #include <xf86drm.h>
20
21 #include "nouveau_winsys.h"
22 #include "nouveau_screen.h"
23 #include "nouveau_context.h"
24 #include "nouveau_fence.h"
25 #include "nouveau_mm.h"
26 #include "nouveau_buffer.h"
27
28 #include <compiler/glsl_types.h>
29
30 /* XXX this should go away */
31 #include "frontend/drm_driver.h"
32
33 /* Even though GPUs might allow addresses with more bits, some engines do not.
34 * Stick with 40 for compatibility.
35 */
36 #define NV_GENERIC_VM_LIMIT_SHIFT 39
37
38 int nouveau_mesa_debug = 0;
39
40 static const char *
nouveau_screen_get_name(struct pipe_screen * pscreen)41 nouveau_screen_get_name(struct pipe_screen *pscreen)
42 {
43 struct nouveau_screen *screen = nouveau_screen(pscreen);
44 return screen->chipset_name;
45 }
46
47 static const char *
nouveau_screen_get_vendor(struct pipe_screen * pscreen)48 nouveau_screen_get_vendor(struct pipe_screen *pscreen)
49 {
50 return "nouveau";
51 }
52
53 static const char *
nouveau_screen_get_device_vendor(struct pipe_screen * pscreen)54 nouveau_screen_get_device_vendor(struct pipe_screen *pscreen)
55 {
56 return "NVIDIA";
57 }
58
59 static uint64_t
nouveau_screen_get_timestamp(struct pipe_screen * pscreen)60 nouveau_screen_get_timestamp(struct pipe_screen *pscreen)
61 {
62 int64_t cpu_time = os_time_get() * 1000;
63
64 /* getparam of PTIMER_TIME takes about x10 as long (several usecs) */
65
66 return cpu_time + nouveau_screen(pscreen)->cpu_gpu_time_delta;
67 }
68
69 static struct disk_cache *
nouveau_screen_get_disk_shader_cache(struct pipe_screen * pscreen)70 nouveau_screen_get_disk_shader_cache(struct pipe_screen *pscreen)
71 {
72 return nouveau_screen(pscreen)->disk_shader_cache;
73 }
74
75 static void
nouveau_screen_fence_ref(struct pipe_screen * pscreen,struct pipe_fence_handle ** ptr,struct pipe_fence_handle * pfence)76 nouveau_screen_fence_ref(struct pipe_screen *pscreen,
77 struct pipe_fence_handle **ptr,
78 struct pipe_fence_handle *pfence)
79 {
80 nouveau_fence_ref(nouveau_fence(pfence), (struct nouveau_fence **)ptr);
81 }
82
83 static bool
nouveau_screen_fence_finish(struct pipe_screen * screen,struct pipe_context * ctx,struct pipe_fence_handle * pfence,uint64_t timeout)84 nouveau_screen_fence_finish(struct pipe_screen *screen,
85 struct pipe_context *ctx,
86 struct pipe_fence_handle *pfence,
87 uint64_t timeout)
88 {
89 if (!timeout)
90 return nouveau_fence_signalled(nouveau_fence(pfence));
91
92 return nouveau_fence_wait(nouveau_fence(pfence), NULL);
93 }
94
95
96 struct nouveau_bo *
nouveau_screen_bo_from_handle(struct pipe_screen * pscreen,struct winsys_handle * whandle,unsigned * out_stride)97 nouveau_screen_bo_from_handle(struct pipe_screen *pscreen,
98 struct winsys_handle *whandle,
99 unsigned *out_stride)
100 {
101 struct nouveau_device *dev = nouveau_screen(pscreen)->device;
102 struct nouveau_bo *bo = NULL;
103 int ret;
104
105 if (whandle->offset != 0) {
106 debug_printf("%s: attempt to import unsupported winsys offset %d\n",
107 __FUNCTION__, whandle->offset);
108 return NULL;
109 }
110
111 if (whandle->type != WINSYS_HANDLE_TYPE_SHARED &&
112 whandle->type != WINSYS_HANDLE_TYPE_FD) {
113 debug_printf("%s: attempt to import unsupported handle type %d\n",
114 __FUNCTION__, whandle->type);
115 return NULL;
116 }
117
118 if (whandle->type == WINSYS_HANDLE_TYPE_SHARED)
119 ret = nouveau_bo_name_ref(dev, whandle->handle, &bo);
120 else
121 ret = nouveau_bo_prime_handle_ref(dev, whandle->handle, &bo);
122
123 if (ret) {
124 debug_printf("%s: ref name 0x%08x failed with %d\n",
125 __FUNCTION__, whandle->handle, ret);
126 return NULL;
127 }
128
129 *out_stride = whandle->stride;
130 return bo;
131 }
132
133
134 bool
nouveau_screen_bo_get_handle(struct pipe_screen * pscreen,struct nouveau_bo * bo,unsigned stride,struct winsys_handle * whandle)135 nouveau_screen_bo_get_handle(struct pipe_screen *pscreen,
136 struct nouveau_bo *bo,
137 unsigned stride,
138 struct winsys_handle *whandle)
139 {
140 whandle->stride = stride;
141
142 if (whandle->type == WINSYS_HANDLE_TYPE_SHARED) {
143 return nouveau_bo_name_get(bo, &whandle->handle) == 0;
144 } else if (whandle->type == WINSYS_HANDLE_TYPE_KMS) {
145 whandle->handle = bo->handle;
146 return true;
147 } else if (whandle->type == WINSYS_HANDLE_TYPE_FD) {
148 return nouveau_bo_set_prime(bo, (int *)&whandle->handle) == 0;
149 } else {
150 return false;
151 }
152 }
153
154 static void
nouveau_disk_cache_create(struct nouveau_screen * screen)155 nouveau_disk_cache_create(struct nouveau_screen *screen)
156 {
157 struct mesa_sha1 ctx;
158 unsigned char sha1[20];
159 char cache_id[20 * 2 + 1];
160 uint64_t driver_flags = 0;
161
162 _mesa_sha1_init(&ctx);
163 if (!disk_cache_get_function_identifier(nouveau_disk_cache_create,
164 &ctx))
165 return;
166
167 _mesa_sha1_final(&ctx, sha1);
168 disk_cache_format_hex_id(cache_id, sha1, 20 * 2);
169
170 if (screen->prefer_nir)
171 driver_flags |= NOUVEAU_SHADER_CACHE_FLAGS_IR_NIR;
172 else
173 driver_flags |= NOUVEAU_SHADER_CACHE_FLAGS_IR_TGSI;
174
175 screen->disk_shader_cache =
176 disk_cache_create(nouveau_screen_get_name(&screen->base),
177 cache_id, driver_flags);
178 }
179
180 static void*
reserve_vma(uintptr_t start,uint64_t reserved_size)181 reserve_vma(uintptr_t start, uint64_t reserved_size)
182 {
183 void *reserved = os_mmap((void*)start, reserved_size, PROT_NONE,
184 MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
185 if (reserved == MAP_FAILED)
186 return NULL;
187 return reserved;
188 }
189
190 int
nouveau_screen_init(struct nouveau_screen * screen,struct nouveau_device * dev)191 nouveau_screen_init(struct nouveau_screen *screen, struct nouveau_device *dev)
192 {
193 struct pipe_screen *pscreen = &screen->base;
194 struct nv04_fifo nv04_data = { .vram = 0xbeef0201, .gart = 0xbeef0202 };
195 struct nvc0_fifo nvc0_data = { };
196 uint64_t time;
197 int size, ret;
198 void *data;
199 union nouveau_bo_config mm_config;
200
201 char *nv_dbg = getenv("NOUVEAU_MESA_DEBUG");
202 if (nv_dbg)
203 nouveau_mesa_debug = atoi(nv_dbg);
204
205 screen->prefer_nir = !debug_get_bool_option("NV50_PROG_USE_TGSI", false);
206
207 screen->force_enable_cl = debug_get_bool_option("NOUVEAU_ENABLE_CL", false);
208 if (screen->force_enable_cl)
209 glsl_type_singleton_init_or_ref();
210
211 screen->disable_fences = debug_get_bool_option("NOUVEAU_DISABLE_FENCES", false);
212
213 /* These must be set before any failure is possible, as the cleanup
214 * paths assume they're responsible for deleting them.
215 */
216 screen->drm = nouveau_drm(&dev->object);
217 screen->device = dev;
218
219 /*
220 * this is initialized to 1 in nouveau_drm_screen_create after screen
221 * is fully constructed and added to the global screen list.
222 */
223 screen->refcount = -1;
224
225 if (dev->chipset < 0xc0) {
226 data = &nv04_data;
227 size = sizeof(nv04_data);
228 } else {
229 data = &nvc0_data;
230 size = sizeof(nvc0_data);
231 }
232
233 bool enable_svm = debug_get_bool_option("NOUVEAU_SVM", false);
234 screen->has_svm = false;
235 /* we only care about HMM with OpenCL enabled */
236 if (dev->chipset > 0x130 && screen->force_enable_cl && enable_svm) {
237 /* Before being able to enable SVM we need to carve out some memory for
238 * driver bo allocations. Let's just base the size on the available VRAM.
239 *
240 * 40 bit is the biggest we care about and for 32 bit systems we don't
241 * want to allocate all of the available memory either.
242 *
243 * Also we align the size we want to reserve to the next POT to make use
244 * of hugepages.
245 */
246 const int vram_shift = util_logbase2_ceil64(dev->vram_size);
247 const int limit_bit =
248 MIN2(sizeof(void*) * 8 - 1, NV_GENERIC_VM_LIMIT_SHIFT);
249 screen->svm_cutout_size =
250 BITFIELD64_BIT(MIN2(sizeof(void*) == 4 ? 26 : NV_GENERIC_VM_LIMIT_SHIFT, vram_shift));
251
252 size_t start = screen->svm_cutout_size;
253 do {
254 screen->svm_cutout = reserve_vma(start, screen->svm_cutout_size);
255 if (!screen->svm_cutout) {
256 start += screen->svm_cutout_size;
257 continue;
258 }
259
260 struct drm_nouveau_svm_init svm_args = {
261 .unmanaged_addr = (uintptr_t)screen->svm_cutout,
262 .unmanaged_size = screen->svm_cutout_size,
263 };
264
265 ret = drmCommandWrite(screen->drm->fd, DRM_NOUVEAU_SVM_INIT,
266 &svm_args, sizeof(svm_args));
267 screen->has_svm = !ret;
268 if (!screen->has_svm)
269 os_munmap(screen->svm_cutout, screen->svm_cutout_size);
270 break;
271 } while ((start + screen->svm_cutout_size) < BITFIELD64_MASK(limit_bit));
272 }
273
274 switch (dev->chipset) {
275 case 0x0ea: /* TK1, GK20A */
276 case 0x12b: /* TX1, GM20B */
277 case 0x13b: /* TX2, GP10B */
278 screen->tegra_sector_layout = true;
279 break;
280 default:
281 /* Xavier's GPU and everything else */
282 screen->tegra_sector_layout = false;
283 break;
284 }
285
286 /*
287 * Set default VRAM domain if not overridden
288 */
289 if (!screen->vram_domain) {
290 if (dev->vram_size > 0)
291 screen->vram_domain = NOUVEAU_BO_VRAM;
292 else
293 screen->vram_domain = NOUVEAU_BO_GART;
294 }
295
296 ret = nouveau_object_new(&dev->object, 0, NOUVEAU_FIFO_CHANNEL_CLASS,
297 data, size, &screen->channel);
298 if (ret)
299 goto err;
300
301 ret = nouveau_client_new(screen->device, &screen->client);
302 if (ret)
303 goto err;
304 ret = nouveau_pushbuf_new(screen->client, screen->channel,
305 4, 512 * 1024, 1,
306 &screen->pushbuf);
307 if (ret)
308 goto err;
309
310 /* getting CPU time first appears to be more accurate */
311 screen->cpu_gpu_time_delta = os_time_get();
312
313 ret = nouveau_getparam(dev, NOUVEAU_GETPARAM_PTIMER_TIME, &time);
314 if (!ret)
315 screen->cpu_gpu_time_delta = time - screen->cpu_gpu_time_delta * 1000;
316
317 snprintf(screen->chipset_name, sizeof(screen->chipset_name), "NV%02X", dev->chipset);
318 pscreen->get_name = nouveau_screen_get_name;
319 pscreen->get_vendor = nouveau_screen_get_vendor;
320 pscreen->get_device_vendor = nouveau_screen_get_device_vendor;
321 pscreen->get_disk_shader_cache = nouveau_screen_get_disk_shader_cache;
322
323 pscreen->get_timestamp = nouveau_screen_get_timestamp;
324
325 pscreen->fence_reference = nouveau_screen_fence_ref;
326 pscreen->fence_finish = nouveau_screen_fence_finish;
327
328 nouveau_disk_cache_create(screen);
329
330 screen->transfer_pushbuf_threshold = 192;
331 screen->lowmem_bindings = PIPE_BIND_GLOBAL; /* gallium limit */
332 screen->vidmem_bindings =
333 PIPE_BIND_RENDER_TARGET | PIPE_BIND_DEPTH_STENCIL |
334 PIPE_BIND_DISPLAY_TARGET | PIPE_BIND_SCANOUT |
335 PIPE_BIND_CURSOR |
336 PIPE_BIND_SAMPLER_VIEW |
337 PIPE_BIND_SHADER_BUFFER | PIPE_BIND_SHADER_IMAGE |
338 PIPE_BIND_COMPUTE_RESOURCE |
339 PIPE_BIND_GLOBAL;
340 screen->sysmem_bindings =
341 PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_STREAM_OUTPUT |
342 PIPE_BIND_COMMAND_ARGS_BUFFER;
343
344 memset(&mm_config, 0, sizeof(mm_config));
345
346 screen->mm_GART = nouveau_mm_create(dev,
347 NOUVEAU_BO_GART | NOUVEAU_BO_MAP,
348 &mm_config);
349 screen->mm_VRAM = nouveau_mm_create(dev, NOUVEAU_BO_VRAM, &mm_config);
350 return 0;
351
352 err:
353 if (screen->svm_cutout)
354 os_munmap(screen->svm_cutout, screen->svm_cutout_size);
355 return ret;
356 }
357
358 void
nouveau_screen_fini(struct nouveau_screen * screen)359 nouveau_screen_fini(struct nouveau_screen *screen)
360 {
361 int fd = screen->drm->fd;
362
363 if (screen->force_enable_cl)
364 glsl_type_singleton_decref();
365 if (screen->has_svm)
366 os_munmap(screen->svm_cutout, screen->svm_cutout_size);
367
368 nouveau_mm_destroy(screen->mm_GART);
369 nouveau_mm_destroy(screen->mm_VRAM);
370
371 nouveau_pushbuf_del(&screen->pushbuf);
372
373 nouveau_client_del(&screen->client);
374 nouveau_object_del(&screen->channel);
375
376 nouveau_device_del(&screen->device);
377 nouveau_drm_del(&screen->drm);
378 close(fd);
379
380 disk_cache_destroy(screen->disk_shader_cache);
381 }
382
383 static void
nouveau_set_debug_callback(struct pipe_context * pipe,const struct util_debug_callback * cb)384 nouveau_set_debug_callback(struct pipe_context *pipe,
385 const struct util_debug_callback *cb)
386 {
387 struct nouveau_context *context = nouveau_context(pipe);
388
389 if (cb)
390 context->debug = *cb;
391 else
392 memset(&context->debug, 0, sizeof(context->debug));
393 }
394
395 void
nouveau_context_init(struct nouveau_context * context)396 nouveau_context_init(struct nouveau_context *context)
397 {
398 context->pipe.set_debug_callback = nouveau_set_debug_callback;
399 }
400