Rev 6104 | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
5078 | serge | 1 | /* |
2 | * Copyright 2009 Jerome Glisse. |
||
3 | * All Rights Reserved. |
||
4 | * |
||
5 | * Permission is hereby granted, free of charge, to any person obtaining a |
||
6 | * copy of this software and associated documentation files (the |
||
7 | * "Software"), to deal in the Software without restriction, including |
||
8 | * without limitation the rights to use, copy, modify, merge, publish, |
||
9 | * distribute, sub license, and/or sell copies of the Software, and to |
||
10 | * permit persons to whom the Software is furnished to do so, subject to |
||
11 | * the following conditions: |
||
12 | * |
||
13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
||
14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
||
15 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL |
||
16 | * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, |
||
17 | * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR |
||
18 | * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE |
||
19 | * USE OR OTHER DEALINGS IN THE SOFTWARE. |
||
20 | * |
||
21 | * The above copyright notice and this permission notice (including the |
||
22 | * next paragraph) shall be included in all copies or substantial portions |
||
23 | * of the Software. |
||
24 | * |
||
25 | */ |
||
26 | /* |
||
27 | * Authors: |
||
28 | * Jerome Glisse |
||
29 | * Thomas Hellstrom |
||
30 | * Dave Airlie |
||
31 | */ |
||
32 | #include |
||
33 | #include |
||
34 | #include |
||
35 | #include |
||
6938 | serge | 36 | #include |
5078 | serge | 37 | #include "radeon.h" |
38 | #include "radeon_trace.h" |
||
39 | |||
40 | |||
41 | int radeon_ttm_init(struct radeon_device *rdev); |
||
42 | void radeon_ttm_fini(struct radeon_device *rdev); |
||
43 | static void radeon_bo_clear_surface_reg(struct radeon_bo *bo); |
||
44 | |||
45 | /* |
||
46 | * To exclude mutual BO access we rely on bo_reserve exclusion, as all |
||
47 | * function are calling it. |
||
48 | */ |
||
49 | |||
50 | static void radeon_update_memory_usage(struct radeon_bo *bo, |
||
51 | unsigned mem_type, int sign) |
||
52 | { |
||
53 | struct radeon_device *rdev = bo->rdev; |
||
54 | u64 size = (u64)bo->tbo.num_pages << PAGE_SHIFT; |
||
55 | |||
56 | switch (mem_type) { |
||
57 | case TTM_PL_TT: |
||
58 | if (sign > 0) |
||
59 | __atomic_add_fetch(&rdev->gtt_usage.counter, size,__ATOMIC_RELAXED); |
||
60 | else |
||
61 | __atomic_sub_fetch(&rdev->gtt_usage.counter, size,__ATOMIC_RELAXED); |
||
62 | break; |
||
63 | case TTM_PL_VRAM: |
||
64 | if (sign > 0) |
||
65 | __atomic_add_fetch(&rdev->vram_usage.counter, size,__ATOMIC_RELAXED); |
||
66 | else |
||
67 | __atomic_sub_fetch(&rdev->vram_usage.counter, size,__ATOMIC_RELAXED ); |
||
68 | break; |
||
69 | } |
||
70 | } |
||
71 | |||
72 | static void radeon_ttm_bo_destroy(struct ttm_buffer_object *tbo) |
||
73 | { |
||
74 | struct radeon_bo *bo; |
||
75 | |||
76 | bo = container_of(tbo, struct radeon_bo, tbo); |
||
77 | |||
78 | radeon_update_memory_usage(bo, bo->tbo.mem.mem_type, -1); |
||
79 | |||
80 | mutex_lock(&bo->rdev->gem.mutex); |
||
81 | list_del_init(&bo->list); |
||
82 | mutex_unlock(&bo->rdev->gem.mutex); |
||
83 | radeon_bo_clear_surface_reg(bo); |
||
84 | WARN_ON(!list_empty(&bo->va)); |
||
85 | drm_gem_object_release(&bo->gem_base); |
||
86 | kfree(bo); |
||
87 | } |
||
88 | |||
89 | bool radeon_ttm_bo_is_radeon_bo(struct ttm_buffer_object *bo) |
||
90 | { |
||
91 | if (bo->destroy == &radeon_ttm_bo_destroy) |
||
92 | return true; |
||
93 | return false; |
||
94 | } |
||
95 | |||
96 | void radeon_ttm_placement_from_domain(struct radeon_bo *rbo, u32 domain) |
||
97 | { |
||
6104 | serge | 98 | u32 c = 0, i; |
5078 | serge | 99 | |
6104 | serge | 100 | rbo->placement.placement = rbo->placements; |
101 | rbo->placement.busy_placement = rbo->placements; |
||
5271 | serge | 102 | if (domain & RADEON_GEM_DOMAIN_VRAM) { |
103 | /* Try placing BOs which don't need CPU access outside of the |
||
104 | * CPU accessible part of VRAM |
||
105 | */ |
||
106 | if ((rbo->flags & RADEON_GEM_NO_CPU_ACCESS) && |
||
107 | rbo->rdev->mc.visible_vram_size < rbo->rdev->mc.real_vram_size) { |
||
108 | rbo->placements[c].fpfn = |
||
109 | rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT; |
||
110 | rbo->placements[c++].flags = TTM_PL_FLAG_WC | |
||
111 | TTM_PL_FLAG_UNCACHED | |
||
6104 | serge | 112 | TTM_PL_FLAG_VRAM; |
5271 | serge | 113 | } |
114 | |||
115 | rbo->placements[c].fpfn = 0; |
||
116 | rbo->placements[c++].flags = TTM_PL_FLAG_WC | |
||
117 | TTM_PL_FLAG_UNCACHED | |
||
118 | TTM_PL_FLAG_VRAM; |
||
119 | } |
||
120 | |||
6104 | serge | 121 | if (domain & RADEON_GEM_DOMAIN_GTT) { |
5078 | serge | 122 | if (rbo->flags & RADEON_GEM_GTT_UC) { |
5271 | serge | 123 | rbo->placements[c].fpfn = 0; |
124 | rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED | |
||
125 | TTM_PL_FLAG_TT; |
||
126 | |||
5078 | serge | 127 | } else if ((rbo->flags & RADEON_GEM_GTT_WC) || |
128 | (rbo->rdev->flags & RADEON_IS_AGP)) { |
||
5271 | serge | 129 | rbo->placements[c].fpfn = 0; |
130 | rbo->placements[c++].flags = TTM_PL_FLAG_WC | |
||
131 | TTM_PL_FLAG_UNCACHED | |
||
5078 | serge | 132 | TTM_PL_FLAG_TT; |
6104 | serge | 133 | } else { |
5271 | serge | 134 | rbo->placements[c].fpfn = 0; |
135 | rbo->placements[c++].flags = TTM_PL_FLAG_CACHED | |
||
136 | TTM_PL_FLAG_TT; |
||
6104 | serge | 137 | } |
138 | } |
||
5271 | serge | 139 | |
6104 | serge | 140 | if (domain & RADEON_GEM_DOMAIN_CPU) { |
5078 | serge | 141 | if (rbo->flags & RADEON_GEM_GTT_UC) { |
5271 | serge | 142 | rbo->placements[c].fpfn = 0; |
143 | rbo->placements[c++].flags = TTM_PL_FLAG_UNCACHED | |
||
144 | TTM_PL_FLAG_SYSTEM; |
||
145 | |||
5078 | serge | 146 | } else if ((rbo->flags & RADEON_GEM_GTT_WC) || |
147 | rbo->rdev->flags & RADEON_IS_AGP) { |
||
5271 | serge | 148 | rbo->placements[c].fpfn = 0; |
149 | rbo->placements[c++].flags = TTM_PL_FLAG_WC | |
||
150 | TTM_PL_FLAG_UNCACHED | |
||
5078 | serge | 151 | TTM_PL_FLAG_SYSTEM; |
6104 | serge | 152 | } else { |
5271 | serge | 153 | rbo->placements[c].fpfn = 0; |
154 | rbo->placements[c++].flags = TTM_PL_FLAG_CACHED | |
||
155 | TTM_PL_FLAG_SYSTEM; |
||
6104 | serge | 156 | } |
157 | } |
||
5271 | serge | 158 | if (!c) { |
159 | rbo->placements[c].fpfn = 0; |
||
160 | rbo->placements[c++].flags = TTM_PL_MASK_CACHING | |
||
161 | TTM_PL_FLAG_SYSTEM; |
||
162 | } |
||
163 | |||
6104 | serge | 164 | rbo->placement.num_placement = c; |
165 | rbo->placement.num_busy_placement = c; |
||
5078 | serge | 166 | |
5271 | serge | 167 | for (i = 0; i < c; ++i) { |
168 | if ((rbo->flags & RADEON_GEM_CPU_ACCESS) && |
||
169 | (rbo->placements[i].flags & TTM_PL_FLAG_VRAM) && |
||
170 | !rbo->placements[i].fpfn) |
||
171 | rbo->placements[i].lpfn = |
||
172 | rbo->rdev->mc.visible_vram_size >> PAGE_SHIFT; |
||
173 | else |
||
174 | rbo->placements[i].lpfn = 0; |
||
175 | } |
||
5078 | serge | 176 | } |
177 | |||
178 | int radeon_bo_create(struct radeon_device *rdev, |
||
5271 | serge | 179 | unsigned long size, int byte_align, bool kernel, |
180 | u32 domain, u32 flags, struct sg_table *sg, |
||
181 | struct reservation_object *resv, |
||
182 | struct radeon_bo **bo_ptr) |
||
5078 | serge | 183 | { |
6104 | serge | 184 | struct radeon_bo *bo; |
5078 | serge | 185 | enum ttm_bo_type type; |
6104 | serge | 186 | unsigned long page_align = roundup(byte_align, PAGE_SIZE) >> PAGE_SHIFT; |
187 | size_t acc_size; |
||
188 | int r; |
||
5078 | serge | 189 | |
190 | size = ALIGN(size, PAGE_SIZE); |
||
191 | |||
192 | if (kernel) { |
||
193 | type = ttm_bo_type_kernel; |
||
194 | } else if (sg) { |
||
195 | type = ttm_bo_type_sg; |
||
196 | } else { |
||
197 | type = ttm_bo_type_device; |
||
198 | } |
||
199 | *bo_ptr = NULL; |
||
200 | |||
201 | acc_size = ttm_bo_dma_acc_size(&rdev->mman.bdev, size, |
||
202 | sizeof(struct radeon_bo)); |
||
203 | |||
6104 | serge | 204 | bo = kzalloc(sizeof(struct radeon_bo), GFP_KERNEL); |
205 | if (bo == NULL) |
||
206 | return -ENOMEM; |
||
207 | r = drm_gem_object_init(rdev->ddev, &bo->gem_base, size); |
||
208 | if (unlikely(r)) { |
||
209 | kfree(bo); |
||
210 | return r; |
||
211 | } |
||
212 | bo->rdev = rdev; |
||
213 | bo->surface_reg = -1; |
||
214 | INIT_LIST_HEAD(&bo->list); |
||
215 | INIT_LIST_HEAD(&bo->va); |
||
216 | bo->initial_domain = domain & (RADEON_GEM_DOMAIN_VRAM | |
||
217 | RADEON_GEM_DOMAIN_GTT | |
||
218 | RADEON_GEM_DOMAIN_CPU); |
||
5078 | serge | 219 | |
220 | bo->flags = flags; |
||
221 | /* PCI GART is always snooped */ |
||
222 | if (!(rdev->flags & RADEON_IS_PCIE)) |
||
223 | bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); |
||
224 | |||
6104 | serge | 225 | /* Write-combined CPU mappings of GTT cause GPU hangs with RV6xx |
226 | * See https://bugs.freedesktop.org/show_bug.cgi?id=91268 |
||
227 | */ |
||
228 | if (rdev->family >= CHIP_RV610 && rdev->family <= CHIP_RV635) |
||
229 | bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); |
||
230 | |||
5271 | serge | 231 | #ifdef CONFIG_X86_32 |
232 | /* XXX: Write-combined CPU mappings of GTT seem broken on 32-bit |
||
233 | * See https://bugs.freedesktop.org/show_bug.cgi?id=84627 |
||
234 | */ |
||
6104 | serge | 235 | bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); |
236 | #elif defined(CONFIG_X86) && !defined(CONFIG_X86_PAT) |
||
237 | /* Don't try to enable write-combining when it can't work, or things |
||
238 | * may be slow |
||
239 | * See https://bugs.freedesktop.org/show_bug.cgi?id=88758 |
||
240 | */ |
||
241 | |||
242 | #warning Please enable CONFIG_MTRR and CONFIG_X86_PAT for better performance \ |
||
243 | thanks to write-combining |
||
244 | |||
245 | if (bo->flags & RADEON_GEM_GTT_WC) |
||
246 | DRM_INFO_ONCE("Please enable CONFIG_MTRR and CONFIG_X86_PAT for " |
||
247 | "better performance thanks to write-combining\n"); |
||
248 | bo->flags &= ~(RADEON_GEM_GTT_WC | RADEON_GEM_GTT_UC); |
||
5271 | serge | 249 | #endif |
5078 | serge | 250 | |
6104 | serge | 251 | radeon_ttm_placement_from_domain(bo, domain); |
5078 | serge | 252 | /* Kernel allocation are uninterruptible */ |
5346 | serge | 253 | down_read(&rdev->pm.mclk_lock); |
5078 | serge | 254 | r = ttm_bo_init(&rdev->mman.bdev, &bo->tbo, size, type, |
255 | &bo->placement, page_align, !kernel, NULL, |
||
5271 | serge | 256 | acc_size, sg, resv, &radeon_ttm_bo_destroy); |
5346 | serge | 257 | up_read(&rdev->pm.mclk_lock); |
5078 | serge | 258 | if (unlikely(r != 0)) { |
259 | return r; |
||
260 | } |
||
261 | *bo_ptr = bo; |
||
262 | |||
263 | trace_radeon_bo_create(bo); |
||
264 | |||
265 | return 0; |
||
266 | } |
||
267 | |||
268 | int radeon_bo_kmap(struct radeon_bo *bo, void **ptr) |
||
269 | { |
||
270 | bool is_iomem; |
||
271 | int r; |
||
272 | |||
273 | if (bo->kptr) { |
||
274 | if (ptr) { |
||
275 | *ptr = bo->kptr; |
||
276 | } |
||
277 | return 0; |
||
278 | } |
||
279 | r = ttm_bo_kmap(&bo->tbo, 0, bo->tbo.num_pages, &bo->kmap); |
||
280 | if (r) { |
||
281 | return r; |
||
282 | } |
||
283 | bo->kptr = ttm_kmap_obj_virtual(&bo->kmap, &is_iomem); |
||
284 | if (ptr) { |
||
285 | *ptr = bo->kptr; |
||
286 | } |
||
287 | radeon_bo_check_tiling(bo, 0, 0); |
||
288 | return 0; |
||
289 | } |
||
290 | |||
291 | void radeon_bo_kunmap(struct radeon_bo *bo) |
||
292 | { |
||
293 | if (bo->kptr == NULL) |
||
294 | return; |
||
295 | bo->kptr = NULL; |
||
296 | radeon_bo_check_tiling(bo, 0, 0); |
||
297 | ttm_bo_kunmap(&bo->kmap); |
||
298 | } |
||
299 | |||
300 | struct radeon_bo *radeon_bo_ref(struct radeon_bo *bo) |
||
301 | { |
||
302 | if (bo == NULL) |
||
303 | return NULL; |
||
304 | |||
305 | ttm_bo_reference(&bo->tbo); |
||
306 | return bo; |
||
307 | } |
||
308 | |||
309 | void radeon_bo_unref(struct radeon_bo **bo) |
||
310 | { |
||
311 | struct ttm_buffer_object *tbo; |
||
312 | struct radeon_device *rdev; |
||
313 | |||
314 | if ((*bo) == NULL) |
||
315 | return; |
||
316 | rdev = (*bo)->rdev; |
||
317 | tbo = &((*bo)->tbo); |
||
318 | ttm_bo_unref(&tbo); |
||
319 | if (tbo == NULL) |
||
320 | *bo = NULL; |
||
321 | } |
||
322 | |||
323 | int radeon_bo_pin_restricted(struct radeon_bo *bo, u32 domain, u64 max_offset, |
||
6104 | serge | 324 | u64 *gpu_addr) |
5078 | serge | 325 | { |
6104 | serge | 326 | int r, i; |
5078 | serge | 327 | |
6104 | serge | 328 | if (bo->pin_count) { |
329 | bo->pin_count++; |
||
330 | if (gpu_addr) |
||
331 | *gpu_addr = radeon_bo_gpu_offset(bo); |
||
5078 | serge | 332 | |
6104 | serge | 333 | if (max_offset != 0) { |
334 | u64 domain_start; |
||
5078 | serge | 335 | |
6104 | serge | 336 | if (domain == RADEON_GEM_DOMAIN_VRAM) |
337 | domain_start = bo->rdev->mc.vram_start; |
||
338 | else |
||
339 | domain_start = bo->rdev->mc.gtt_start; |
||
340 | WARN_ON_ONCE(max_offset < |
||
341 | (radeon_bo_gpu_offset(bo) - domain_start)); |
||
342 | } |
||
5078 | serge | 343 | |
6104 | serge | 344 | return 0; |
345 | } |
||
346 | radeon_ttm_placement_from_domain(bo, domain); |
||
5271 | serge | 347 | for (i = 0; i < bo->placement.num_placement; i++) { |
6104 | serge | 348 | /* force to pin into visible video ram */ |
5271 | serge | 349 | if ((bo->placements[i].flags & TTM_PL_FLAG_VRAM) && |
350 | !(bo->flags & RADEON_GEM_NO_CPU_ACCESS) && |
||
351 | (!max_offset || max_offset > bo->rdev->mc.visible_vram_size)) |
||
352 | bo->placements[i].lpfn = |
||
353 | bo->rdev->mc.visible_vram_size >> PAGE_SHIFT; |
||
354 | else |
||
355 | bo->placements[i].lpfn = max_offset >> PAGE_SHIFT; |
||
356 | |||
357 | bo->placements[i].flags |= TTM_PL_FLAG_NO_EVICT; |
||
6104 | serge | 358 | } |
5078 | serge | 359 | |
360 | r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); |
||
6104 | serge | 361 | if (likely(r == 0)) { |
362 | bo->pin_count = 1; |
||
363 | if (gpu_addr != NULL) |
||
364 | *gpu_addr = radeon_bo_gpu_offset(bo); |
||
5078 | serge | 365 | if (domain == RADEON_GEM_DOMAIN_VRAM) |
366 | bo->rdev->vram_pin_size += radeon_bo_size(bo); |
||
367 | else |
||
368 | bo->rdev->gart_pin_size += radeon_bo_size(bo); |
||
369 | } else { |
||
370 | dev_err(bo->rdev->dev, "%p pin failed\n", bo); |
||
6104 | serge | 371 | } |
372 | return r; |
||
5078 | serge | 373 | } |
374 | |||
375 | int radeon_bo_pin(struct radeon_bo *bo, u32 domain, u64 *gpu_addr) |
||
376 | { |
||
6104 | serge | 377 | return radeon_bo_pin_restricted(bo, domain, 0, gpu_addr); |
5078 | serge | 378 | } |
379 | |||
380 | int radeon_bo_unpin(struct radeon_bo *bo) |
||
381 | { |
||
6104 | serge | 382 | int r, i; |
5078 | serge | 383 | |
6104 | serge | 384 | if (!bo->pin_count) { |
385 | dev_warn(bo->rdev->dev, "%p unpin not necessary\n", bo); |
||
386 | return 0; |
||
387 | } |
||
388 | bo->pin_count--; |
||
389 | if (bo->pin_count) |
||
390 | return 0; |
||
5271 | serge | 391 | for (i = 0; i < bo->placement.num_placement; i++) { |
392 | bo->placements[i].lpfn = 0; |
||
393 | bo->placements[i].flags &= ~TTM_PL_FLAG_NO_EVICT; |
||
394 | } |
||
5078 | serge | 395 | r = ttm_bo_validate(&bo->tbo, &bo->placement, false, false); |
396 | if (likely(r == 0)) { |
||
397 | if (bo->tbo.mem.mem_type == TTM_PL_VRAM) |
||
398 | bo->rdev->vram_pin_size -= radeon_bo_size(bo); |
||
399 | else |
||
400 | bo->rdev->gart_pin_size -= radeon_bo_size(bo); |
||
401 | } else { |
||
402 | dev_err(bo->rdev->dev, "%p validate failed for unpin\n", bo); |
||
403 | } |
||
404 | return r; |
||
405 | } |
||
406 | |||
407 | int radeon_bo_init(struct radeon_device *rdev) |
||
408 | { |
||
409 | /* Add an MTRR for the VRAM */ |
||
410 | DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n", |
||
411 | rdev->mc.mc_vram_size >> 20, |
||
412 | (unsigned long long)rdev->mc.aper_size >> 20); |
||
413 | DRM_INFO("RAM width %dbits %cDR\n", |
||
414 | rdev->mc.vram_width, rdev->mc.vram_is_ddr ? 'D' : 'S'); |
||
415 | return radeon_ttm_init(rdev); |
||
416 | } |
||
417 | |||
418 | void radeon_bo_fini(struct radeon_device *rdev) |
||
419 | { |
||
420 | // radeon_ttm_fini(rdev); |
||
421 | // arch_phys_wc_del(rdev->mc.vram_mtrr); |
||
422 | } |
||
423 | |||
424 | /* Returns how many bytes TTM can move per IB. |
||
425 | */ |
||
426 | static u64 radeon_bo_get_threshold_for_moves(struct radeon_device *rdev) |
||
427 | { |
||
428 | u64 real_vram_size = rdev->mc.real_vram_size; |
||
429 | u64 vram_usage = atomic64_read(&rdev->vram_usage); |
||
430 | |||
431 | /* This function is based on the current VRAM usage. |
||
432 | * |
||
433 | * - If all of VRAM is free, allow relocating the number of bytes that |
||
434 | * is equal to 1/4 of the size of VRAM for this IB. |
||
435 | |||
436 | * - If more than one half of VRAM is occupied, only allow relocating |
||
437 | * 1 MB of data for this IB. |
||
438 | * |
||
439 | * - From 0 to one half of used VRAM, the threshold decreases |
||
440 | * linearly. |
||
441 | * __________________ |
||
442 | * 1/4 of -|\ | |
||
443 | * VRAM | \ | |
||
444 | * | \ | |
||
445 | * | \ | |
||
446 | * | \ | |
||
447 | * | \ | |
||
448 | * | \ | |
||
449 | * | \________|1 MB |
||
450 | * |----------------| |
||
451 | * VRAM 0 % 100 % |
||
452 | * used used |
||
453 | * |
||
454 | * Note: It's a threshold, not a limit. The threshold must be crossed |
||
455 | * for buffer relocations to stop, so any buffer of an arbitrary size |
||
456 | * can be moved as long as the threshold isn't crossed before |
||
457 | * the relocation takes place. We don't want to disable buffer |
||
458 | * relocations completely. |
||
459 | * |
||
460 | * The idea is that buffers should be placed in VRAM at creation time |
||
461 | * and TTM should only do a minimum number of relocations during |
||
462 | * command submission. In practice, you need to submit at least |
||
463 | * a dozen IBs to move all buffers to VRAM if they are in GTT. |
||
464 | * |
||
465 | * Also, things can get pretty crazy under memory pressure and actual |
||
466 | * VRAM usage can change a lot, so playing safe even at 50% does |
||
467 | * consistently increase performance. |
||
468 | */ |
||
469 | |||
470 | u64 half_vram = real_vram_size >> 1; |
||
471 | u64 half_free_vram = vram_usage >= half_vram ? 0 : half_vram - vram_usage; |
||
472 | u64 bytes_moved_threshold = half_free_vram >> 1; |
||
473 | return max(bytes_moved_threshold, 1024*1024ull); |
||
474 | } |
||
475 | |||
476 | int radeon_bo_list_validate(struct radeon_device *rdev, |
||
477 | struct ww_acquire_ctx *ticket, |
||
478 | struct list_head *head, int ring) |
||
479 | { |
||
5271 | serge | 480 | struct radeon_bo_list *lobj; |
481 | struct list_head duplicates; |
||
5078 | serge | 482 | int r; |
483 | u64 bytes_moved = 0, initial_bytes_moved; |
||
484 | u64 bytes_moved_threshold = radeon_bo_get_threshold_for_moves(rdev); |
||
485 | |||
5271 | serge | 486 | INIT_LIST_HEAD(&duplicates); |
487 | r = ttm_eu_reserve_buffers(ticket, head, true, &duplicates); |
||
5078 | serge | 488 | if (unlikely(r != 0)) { |
489 | return r; |
||
490 | } |
||
491 | |||
492 | list_for_each_entry(lobj, head, tv.head) { |
||
5271 | serge | 493 | struct radeon_bo *bo = lobj->robj; |
5078 | serge | 494 | if (!bo->pin_count) { |
495 | u32 domain = lobj->prefered_domains; |
||
5271 | serge | 496 | u32 allowed = lobj->allowed_domains; |
5078 | serge | 497 | u32 current_domain = |
498 | radeon_mem_type_to_domain(bo->tbo.mem.mem_type); |
||
499 | |||
500 | /* Check if this buffer will be moved and don't move it |
||
501 | * if we have moved too many buffers for this IB already. |
||
502 | * |
||
503 | * Note that this allows moving at least one buffer of |
||
504 | * any size, because it doesn't take the current "bo" |
||
505 | * into account. We don't want to disallow buffer moves |
||
506 | * completely. |
||
507 | */ |
||
5271 | serge | 508 | if ((allowed & current_domain) != 0 && |
5078 | serge | 509 | (domain & current_domain) == 0 && /* will be moved */ |
510 | bytes_moved > bytes_moved_threshold) { |
||
511 | /* don't move it */ |
||
512 | domain = current_domain; |
||
513 | } |
||
514 | |||
515 | retry: |
||
516 | radeon_ttm_placement_from_domain(bo, domain); |
||
517 | if (ring == R600_RING_TYPE_UVD_INDEX) |
||
5271 | serge | 518 | radeon_uvd_force_into_uvd_segment(bo, allowed); |
5078 | serge | 519 | |
520 | initial_bytes_moved = atomic64_read(&rdev->num_bytes_moved); |
||
521 | r = ttm_bo_validate(&bo->tbo, &bo->placement, true, false); |
||
522 | bytes_moved += atomic64_read(&rdev->num_bytes_moved) - |
||
523 | initial_bytes_moved; |
||
524 | |||
525 | if (unlikely(r)) { |
||
526 | if (r != -ERESTARTSYS && |
||
527 | domain != lobj->allowed_domains) { |
||
528 | domain = lobj->allowed_domains; |
||
529 | goto retry; |
||
530 | } |
||
531 | ttm_eu_backoff_reservation(ticket, head); |
||
532 | return r; |
||
533 | } |
||
534 | } |
||
535 | lobj->gpu_offset = radeon_bo_gpu_offset(bo); |
||
536 | lobj->tiling_flags = bo->tiling_flags; |
||
537 | } |
||
5271 | serge | 538 | |
539 | list_for_each_entry(lobj, &duplicates, tv.head) { |
||
540 | lobj->gpu_offset = radeon_bo_gpu_offset(lobj->robj); |
||
541 | lobj->tiling_flags = lobj->robj->tiling_flags; |
||
542 | } |
||
543 | |||
5078 | serge | 544 | return 0; |
545 | } |
||
546 | |||
547 | int radeon_bo_get_surface_reg(struct radeon_bo *bo) |
||
548 | { |
||
549 | struct radeon_device *rdev = bo->rdev; |
||
550 | struct radeon_surface_reg *reg; |
||
551 | struct radeon_bo *old_object; |
||
552 | int steal; |
||
553 | int i; |
||
554 | |||
555 | lockdep_assert_held(&bo->tbo.resv->lock.base); |
||
556 | |||
557 | if (!bo->tiling_flags) |
||
558 | return 0; |
||
559 | |||
560 | if (bo->surface_reg >= 0) { |
||
561 | reg = &rdev->surface_regs[bo->surface_reg]; |
||
562 | i = bo->surface_reg; |
||
563 | goto out; |
||
564 | } |
||
565 | |||
566 | steal = -1; |
||
567 | for (i = 0; i < RADEON_GEM_MAX_SURFACES; i++) { |
||
568 | |||
569 | reg = &rdev->surface_regs[i]; |
||
570 | if (!reg->bo) |
||
571 | break; |
||
572 | |||
573 | old_object = reg->bo; |
||
574 | if (old_object->pin_count == 0) |
||
575 | steal = i; |
||
576 | } |
||
577 | |||
578 | /* if we are all out */ |
||
579 | if (i == RADEON_GEM_MAX_SURFACES) { |
||
580 | if (steal == -1) |
||
581 | return -ENOMEM; |
||
582 | /* find someone with a surface reg and nuke their BO */ |
||
583 | reg = &rdev->surface_regs[steal]; |
||
584 | old_object = reg->bo; |
||
585 | /* blow away the mapping */ |
||
586 | DRM_DEBUG("stealing surface reg %d from %p\n", steal, old_object); |
||
587 | ttm_bo_unmap_virtual(&old_object->tbo); |
||
588 | old_object->surface_reg = -1; |
||
589 | i = steal; |
||
590 | } |
||
591 | |||
592 | bo->surface_reg = i; |
||
593 | reg->bo = bo; |
||
594 | |||
595 | out: |
||
596 | radeon_set_surface_reg(rdev, i, bo->tiling_flags, bo->pitch, |
||
597 | bo->tbo.mem.start << PAGE_SHIFT, |
||
598 | bo->tbo.num_pages << PAGE_SHIFT); |
||
599 | return 0; |
||
600 | } |
||
601 | |||
602 | static void radeon_bo_clear_surface_reg(struct radeon_bo *bo) |
||
603 | { |
||
604 | struct radeon_device *rdev = bo->rdev; |
||
605 | struct radeon_surface_reg *reg; |
||
606 | |||
607 | if (bo->surface_reg == -1) |
||
608 | return; |
||
609 | |||
610 | reg = &rdev->surface_regs[bo->surface_reg]; |
||
611 | radeon_clear_surface_reg(rdev, bo->surface_reg); |
||
612 | |||
613 | reg->bo = NULL; |
||
614 | bo->surface_reg = -1; |
||
615 | } |
||
616 | |||
617 | int radeon_bo_set_tiling_flags(struct radeon_bo *bo, |
||
618 | uint32_t tiling_flags, uint32_t pitch) |
||
619 | { |
||
620 | struct radeon_device *rdev = bo->rdev; |
||
621 | int r; |
||
622 | |||
623 | if (rdev->family >= CHIP_CEDAR) { |
||
624 | unsigned bankw, bankh, mtaspect, tilesplit, stilesplit; |
||
625 | |||
626 | bankw = (tiling_flags >> RADEON_TILING_EG_BANKW_SHIFT) & RADEON_TILING_EG_BANKW_MASK; |
||
627 | bankh = (tiling_flags >> RADEON_TILING_EG_BANKH_SHIFT) & RADEON_TILING_EG_BANKH_MASK; |
||
628 | mtaspect = (tiling_flags >> RADEON_TILING_EG_MACRO_TILE_ASPECT_SHIFT) & RADEON_TILING_EG_MACRO_TILE_ASPECT_MASK; |
||
629 | tilesplit = (tiling_flags >> RADEON_TILING_EG_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_TILE_SPLIT_MASK; |
||
630 | stilesplit = (tiling_flags >> RADEON_TILING_EG_STENCIL_TILE_SPLIT_SHIFT) & RADEON_TILING_EG_STENCIL_TILE_SPLIT_MASK; |
||
631 | switch (bankw) { |
||
632 | case 0: |
||
633 | case 1: |
||
634 | case 2: |
||
635 | case 4: |
||
636 | case 8: |
||
637 | break; |
||
638 | default: |
||
639 | return -EINVAL; |
||
640 | } |
||
641 | switch (bankh) { |
||
642 | case 0: |
||
643 | case 1: |
||
644 | case 2: |
||
645 | case 4: |
||
646 | case 8: |
||
647 | break; |
||
648 | default: |
||
649 | return -EINVAL; |
||
650 | } |
||
651 | switch (mtaspect) { |
||
652 | case 0: |
||
653 | case 1: |
||
654 | case 2: |
||
655 | case 4: |
||
656 | case 8: |
||
657 | break; |
||
658 | default: |
||
659 | return -EINVAL; |
||
660 | } |
||
661 | if (tilesplit > 6) { |
||
662 | return -EINVAL; |
||
663 | } |
||
664 | if (stilesplit > 6) { |
||
665 | return -EINVAL; |
||
666 | } |
||
667 | } |
||
668 | r = radeon_bo_reserve(bo, false); |
||
669 | if (unlikely(r != 0)) |
||
670 | return r; |
||
671 | bo->tiling_flags = tiling_flags; |
||
672 | bo->pitch = pitch; |
||
673 | radeon_bo_unreserve(bo); |
||
674 | return 0; |
||
675 | } |
||
676 | |||
677 | void radeon_bo_get_tiling_flags(struct radeon_bo *bo, |
||
678 | uint32_t *tiling_flags, |
||
679 | uint32_t *pitch) |
||
680 | { |
||
681 | lockdep_assert_held(&bo->tbo.resv->lock.base); |
||
682 | |||
683 | if (tiling_flags) |
||
684 | *tiling_flags = bo->tiling_flags; |
||
685 | if (pitch) |
||
686 | *pitch = bo->pitch; |
||
687 | } |
||
688 | |||
689 | int radeon_bo_check_tiling(struct radeon_bo *bo, bool has_moved, |
||
690 | bool force_drop) |
||
691 | { |
||
692 | if (!force_drop) |
||
693 | lockdep_assert_held(&bo->tbo.resv->lock.base); |
||
694 | |||
695 | if (!(bo->tiling_flags & RADEON_TILING_SURFACE)) |
||
696 | return 0; |
||
697 | |||
698 | if (force_drop) { |
||
699 | radeon_bo_clear_surface_reg(bo); |
||
700 | return 0; |
||
701 | } |
||
702 | |||
703 | if (bo->tbo.mem.mem_type != TTM_PL_VRAM) { |
||
704 | if (!has_moved) |
||
705 | return 0; |
||
706 | |||
707 | if (bo->surface_reg >= 0) |
||
708 | radeon_bo_clear_surface_reg(bo); |
||
709 | return 0; |
||
710 | } |
||
711 | |||
712 | if ((bo->surface_reg >= 0) && !has_moved) |
||
713 | return 0; |
||
714 | |||
715 | return radeon_bo_get_surface_reg(bo); |
||
716 | } |
||
717 | |||
718 | void radeon_bo_move_notify(struct ttm_buffer_object *bo, |
||
719 | struct ttm_mem_reg *new_mem) |
||
720 | { |
||
721 | struct radeon_bo *rbo; |
||
722 | |||
723 | if (!radeon_ttm_bo_is_radeon_bo(bo)) |
||
724 | return; |
||
725 | |||
726 | rbo = container_of(bo, struct radeon_bo, tbo); |
||
727 | radeon_bo_check_tiling(rbo, 0, 1); |
||
728 | radeon_vm_bo_invalidate(rbo->rdev, rbo); |
||
729 | |||
730 | /* update statistics */ |
||
731 | if (!new_mem) |
||
732 | return; |
||
733 | |||
734 | radeon_update_memory_usage(rbo, bo->mem.mem_type, -1); |
||
735 | radeon_update_memory_usage(rbo, new_mem->mem_type, 1); |
||
736 | } |
||
737 | int radeon_bo_wait(struct radeon_bo *bo, u32 *mem_type, bool no_wait) |
||
738 | { |
||
739 | int r; |
||
740 | |||
741 | r = ttm_bo_reserve(&bo->tbo, true, no_wait, false, NULL); |
||
742 | if (unlikely(r != 0)) |
||
743 | return r; |
||
744 | if (mem_type) |
||
745 | *mem_type = bo->tbo.mem.mem_type; |
||
5271 | serge | 746 | |
6104 | serge | 747 | r = ttm_bo_wait(&bo->tbo, true, true, no_wait); |
5078 | serge | 748 | ttm_bo_unreserve(&bo->tbo); |
749 | return r; |
||
750 | } |
||
5271 | serge | 751 | |
752 | /** |
||
753 | * radeon_bo_fence - add fence to buffer object |
||
754 | * |
||
755 | * @bo: buffer object in question |
||
756 | * @fence: fence to add |
||
757 | * @shared: true if fence should be added shared |
||
758 | * |
||
759 | */ |
||
760 | void radeon_bo_fence(struct radeon_bo *bo, struct radeon_fence *fence, |
||
761 | bool shared) |
||
762 | { |
||
763 | struct reservation_object *resv = bo->tbo.resv; |
||
764 | |||
765 | if (shared) |
||
766 | reservation_object_add_shared_fence(resv, &fence->base); |
||
767 | else |
||
768 | reservation_object_add_excl_fence(resv, &fence->base); |
||
769 | }><>><>>>> |