1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
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 * Authors: Christian König <christian.koenig@amd.com>
26 */
27
28 #include <linux/firmware.h>
29
30 #include "amdgpu.h"
31 #include "amdgpu_vce.h"
32 #include "vid.h"
33 #include "vce/vce_3_0_d.h"
34 #include "vce/vce_3_0_sh_mask.h"
35 #include "oss/oss_3_0_d.h"
36 #include "oss/oss_3_0_sh_mask.h"
37 #include "gca/gfx_8_0_d.h"
38 #include "smu/smu_7_1_2_d.h"
39 #include "smu/smu_7_1_2_sh_mask.h"
40 #include "gca/gfx_8_0_sh_mask.h"
41 #include "ivsrcid/ivsrcid_vislands30.h"
42
43
44 #define GRBM_GFX_INDEX__VCE_INSTANCE__SHIFT 0x04
45 #define GRBM_GFX_INDEX__VCE_INSTANCE_MASK 0x10
46 #define GRBM_GFX_INDEX__VCE_ALL_PIPE 0x07
47
48 #define mmVCE_LMI_VCPU_CACHE_40BIT_BAR0 0x8616
49 #define mmVCE_LMI_VCPU_CACHE_40BIT_BAR1 0x8617
50 #define mmVCE_LMI_VCPU_CACHE_40BIT_BAR2 0x8618
51 #define mmGRBM_GFX_INDEX_DEFAULT 0xE0000000
52
53 #define VCE_STATUS_VCPU_REPORT_FW_LOADED_MASK 0x02
54
55 #define VCE_V3_0_FW_SIZE (384 * 1024)
56 #define VCE_V3_0_STACK_SIZE (64 * 1024)
57 #define VCE_V3_0_DATA_SIZE ((16 * 1024 * AMDGPU_MAX_VCE_HANDLES) + (52 * 1024))
58
59 #define FW_52_8_3 ((52 << 24) | (8 << 16) | (3 << 8))
60
61 #define GET_VCE_INSTANCE(i) ((i) << GRBM_GFX_INDEX__VCE_INSTANCE__SHIFT \
62 | GRBM_GFX_INDEX__VCE_ALL_PIPE)
63
64 static void vce_v3_0_mc_resume(struct amdgpu_device *adev, int idx);
65 static void vce_v3_0_set_ring_funcs(struct amdgpu_device *adev);
66 static void vce_v3_0_set_irq_funcs(struct amdgpu_device *adev);
67 static int vce_v3_0_wait_for_idle(struct amdgpu_ip_block *ip_block);
68 static int vce_v3_0_set_clockgating_state(struct amdgpu_ip_block *ip_block,
69 enum amd_clockgating_state state);
70 /**
71 * vce_v3_0_ring_get_rptr - get read pointer
72 *
73 * @ring: amdgpu_ring pointer
74 *
75 * Returns the current hardware read pointer
76 */
vce_v3_0_ring_get_rptr(struct amdgpu_ring * ring)77 static uint64_t vce_v3_0_ring_get_rptr(struct amdgpu_ring *ring)
78 {
79 struct amdgpu_device *adev = ring->adev;
80 u32 v;
81
82 mutex_lock(&adev->grbm_idx_mutex);
83 if (adev->vce.harvest_config == 0 ||
84 adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE1)
85 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(0));
86 else if (adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE0)
87 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(1));
88
89 if (ring->me == 0)
90 v = RREG32(mmVCE_RB_RPTR);
91 else if (ring->me == 1)
92 v = RREG32(mmVCE_RB_RPTR2);
93 else
94 v = RREG32(mmVCE_RB_RPTR3);
95
96 WREG32(mmGRBM_GFX_INDEX, mmGRBM_GFX_INDEX_DEFAULT);
97 mutex_unlock(&adev->grbm_idx_mutex);
98
99 return v;
100 }
101
102 /**
103 * vce_v3_0_ring_get_wptr - get write pointer
104 *
105 * @ring: amdgpu_ring pointer
106 *
107 * Returns the current hardware write pointer
108 */
vce_v3_0_ring_get_wptr(struct amdgpu_ring * ring)109 static uint64_t vce_v3_0_ring_get_wptr(struct amdgpu_ring *ring)
110 {
111 struct amdgpu_device *adev = ring->adev;
112 u32 v;
113
114 mutex_lock(&adev->grbm_idx_mutex);
115 if (adev->vce.harvest_config == 0 ||
116 adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE1)
117 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(0));
118 else if (adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE0)
119 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(1));
120
121 if (ring->me == 0)
122 v = RREG32(mmVCE_RB_WPTR);
123 else if (ring->me == 1)
124 v = RREG32(mmVCE_RB_WPTR2);
125 else
126 v = RREG32(mmVCE_RB_WPTR3);
127
128 WREG32(mmGRBM_GFX_INDEX, mmGRBM_GFX_INDEX_DEFAULT);
129 mutex_unlock(&adev->grbm_idx_mutex);
130
131 return v;
132 }
133
134 /**
135 * vce_v3_0_ring_set_wptr - set write pointer
136 *
137 * @ring: amdgpu_ring pointer
138 *
139 * Commits the write pointer to the hardware
140 */
vce_v3_0_ring_set_wptr(struct amdgpu_ring * ring)141 static void vce_v3_0_ring_set_wptr(struct amdgpu_ring *ring)
142 {
143 struct amdgpu_device *adev = ring->adev;
144
145 mutex_lock(&adev->grbm_idx_mutex);
146 if (adev->vce.harvest_config == 0 ||
147 adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE1)
148 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(0));
149 else if (adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE0)
150 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(1));
151
152 if (ring->me == 0)
153 WREG32(mmVCE_RB_WPTR, lower_32_bits(ring->wptr));
154 else if (ring->me == 1)
155 WREG32(mmVCE_RB_WPTR2, lower_32_bits(ring->wptr));
156 else
157 WREG32(mmVCE_RB_WPTR3, lower_32_bits(ring->wptr));
158
159 WREG32(mmGRBM_GFX_INDEX, mmGRBM_GFX_INDEX_DEFAULT);
160 mutex_unlock(&adev->grbm_idx_mutex);
161 }
162
vce_v3_0_override_vce_clock_gating(struct amdgpu_device * adev,bool override)163 static void vce_v3_0_override_vce_clock_gating(struct amdgpu_device *adev, bool override)
164 {
165 WREG32_FIELD(VCE_RB_ARB_CTRL, VCE_CGTT_OVERRIDE, override ? 1 : 0);
166 }
167
vce_v3_0_set_vce_sw_clock_gating(struct amdgpu_device * adev,bool gated)168 static void vce_v3_0_set_vce_sw_clock_gating(struct amdgpu_device *adev,
169 bool gated)
170 {
171 u32 data;
172
173 /* Set Override to disable Clock Gating */
174 vce_v3_0_override_vce_clock_gating(adev, true);
175
176 /* This function enables MGCG which is controlled by firmware.
177 With the clocks in the gated state the core is still
178 accessible but the firmware will throttle the clocks on the
179 fly as necessary.
180 */
181 if (!gated) {
182 data = RREG32(mmVCE_CLOCK_GATING_B);
183 data |= 0x1ff;
184 data &= ~0xef0000;
185 WREG32(mmVCE_CLOCK_GATING_B, data);
186
187 data = RREG32(mmVCE_UENC_CLOCK_GATING);
188 data |= 0x3ff000;
189 data &= ~0xffc00000;
190 WREG32(mmVCE_UENC_CLOCK_GATING, data);
191
192 data = RREG32(mmVCE_UENC_CLOCK_GATING_2);
193 data |= 0x2;
194 data &= ~0x00010000;
195 WREG32(mmVCE_UENC_CLOCK_GATING_2, data);
196
197 data = RREG32(mmVCE_UENC_REG_CLOCK_GATING);
198 data |= 0x37f;
199 WREG32(mmVCE_UENC_REG_CLOCK_GATING, data);
200
201 data = RREG32(mmVCE_UENC_DMA_DCLK_CTRL);
202 data |= VCE_UENC_DMA_DCLK_CTRL__WRDMCLK_FORCEON_MASK |
203 VCE_UENC_DMA_DCLK_CTRL__RDDMCLK_FORCEON_MASK |
204 VCE_UENC_DMA_DCLK_CTRL__REGCLK_FORCEON_MASK |
205 0x8;
206 WREG32(mmVCE_UENC_DMA_DCLK_CTRL, data);
207 } else {
208 data = RREG32(mmVCE_CLOCK_GATING_B);
209 data &= ~0x80010;
210 data |= 0xe70008;
211 WREG32(mmVCE_CLOCK_GATING_B, data);
212
213 data = RREG32(mmVCE_UENC_CLOCK_GATING);
214 data |= 0xffc00000;
215 WREG32(mmVCE_UENC_CLOCK_GATING, data);
216
217 data = RREG32(mmVCE_UENC_CLOCK_GATING_2);
218 data |= 0x10000;
219 WREG32(mmVCE_UENC_CLOCK_GATING_2, data);
220
221 data = RREG32(mmVCE_UENC_REG_CLOCK_GATING);
222 data &= ~0x3ff;
223 WREG32(mmVCE_UENC_REG_CLOCK_GATING, data);
224
225 data = RREG32(mmVCE_UENC_DMA_DCLK_CTRL);
226 data &= ~(VCE_UENC_DMA_DCLK_CTRL__WRDMCLK_FORCEON_MASK |
227 VCE_UENC_DMA_DCLK_CTRL__RDDMCLK_FORCEON_MASK |
228 VCE_UENC_DMA_DCLK_CTRL__REGCLK_FORCEON_MASK |
229 0x8);
230 WREG32(mmVCE_UENC_DMA_DCLK_CTRL, data);
231 }
232 vce_v3_0_override_vce_clock_gating(adev, false);
233 }
234
vce_v3_0_firmware_loaded(struct amdgpu_device * adev)235 static int vce_v3_0_firmware_loaded(struct amdgpu_device *adev)
236 {
237 int i, j;
238
239 for (i = 0; i < 10; ++i) {
240 for (j = 0; j < 100; ++j) {
241 uint32_t status = RREG32(mmVCE_STATUS);
242
243 if (status & VCE_STATUS_VCPU_REPORT_FW_LOADED_MASK)
244 return 0;
245 mdelay(10);
246 }
247
248 DRM_ERROR("VCE not responding, trying to reset the ECPU!!!\n");
249 WREG32_FIELD(VCE_SOFT_RESET, ECPU_SOFT_RESET, 1);
250 mdelay(10);
251 WREG32_FIELD(VCE_SOFT_RESET, ECPU_SOFT_RESET, 0);
252 mdelay(10);
253 }
254
255 return -ETIMEDOUT;
256 }
257
258 /**
259 * vce_v3_0_start - start VCE block
260 *
261 * @adev: amdgpu_device pointer
262 *
263 * Setup and start the VCE block
264 */
vce_v3_0_start(struct amdgpu_device * adev)265 static int vce_v3_0_start(struct amdgpu_device *adev)
266 {
267 struct amdgpu_ring *ring;
268 int idx, r;
269
270 mutex_lock(&adev->grbm_idx_mutex);
271 for (idx = 0; idx < 2; ++idx) {
272 if (adev->vce.harvest_config & (1 << idx))
273 continue;
274
275 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(idx));
276
277 /* Program instance 0 reg space for two instances or instance 0 case
278 program instance 1 reg space for only instance 1 available case */
279 if (idx != 1 || adev->vce.harvest_config == AMDGPU_VCE_HARVEST_VCE0) {
280 ring = &adev->vce.ring[0];
281 WREG32(mmVCE_RB_RPTR, lower_32_bits(ring->wptr));
282 WREG32(mmVCE_RB_WPTR, lower_32_bits(ring->wptr));
283 WREG32(mmVCE_RB_BASE_LO, ring->gpu_addr);
284 WREG32(mmVCE_RB_BASE_HI, upper_32_bits(ring->gpu_addr));
285 WREG32(mmVCE_RB_SIZE, ring->ring_size / 4);
286
287 ring = &adev->vce.ring[1];
288 WREG32(mmVCE_RB_RPTR2, lower_32_bits(ring->wptr));
289 WREG32(mmVCE_RB_WPTR2, lower_32_bits(ring->wptr));
290 WREG32(mmVCE_RB_BASE_LO2, ring->gpu_addr);
291 WREG32(mmVCE_RB_BASE_HI2, upper_32_bits(ring->gpu_addr));
292 WREG32(mmVCE_RB_SIZE2, ring->ring_size / 4);
293
294 ring = &adev->vce.ring[2];
295 WREG32(mmVCE_RB_RPTR3, lower_32_bits(ring->wptr));
296 WREG32(mmVCE_RB_WPTR3, lower_32_bits(ring->wptr));
297 WREG32(mmVCE_RB_BASE_LO3, ring->gpu_addr);
298 WREG32(mmVCE_RB_BASE_HI3, upper_32_bits(ring->gpu_addr));
299 WREG32(mmVCE_RB_SIZE3, ring->ring_size / 4);
300 }
301
302 vce_v3_0_mc_resume(adev, idx);
303 WREG32_FIELD(VCE_STATUS, JOB_BUSY, 1);
304
305 if (adev->asic_type >= CHIP_STONEY)
306 WREG32_P(mmVCE_VCPU_CNTL, 1, ~0x200001);
307 else
308 WREG32_FIELD(VCE_VCPU_CNTL, CLK_EN, 1);
309
310 WREG32_FIELD(VCE_SOFT_RESET, ECPU_SOFT_RESET, 0);
311 mdelay(100);
312
313 r = vce_v3_0_firmware_loaded(adev);
314
315 /* clear BUSY flag */
316 WREG32_FIELD(VCE_STATUS, JOB_BUSY, 0);
317
318 if (r) {
319 DRM_ERROR("VCE not responding, giving up!!!\n");
320 mutex_unlock(&adev->grbm_idx_mutex);
321 return r;
322 }
323 }
324
325 WREG32(mmGRBM_GFX_INDEX, mmGRBM_GFX_INDEX_DEFAULT);
326 mutex_unlock(&adev->grbm_idx_mutex);
327
328 return 0;
329 }
330
vce_v3_0_stop(struct amdgpu_device * adev)331 static int vce_v3_0_stop(struct amdgpu_device *adev)
332 {
333 int idx;
334
335 mutex_lock(&adev->grbm_idx_mutex);
336 for (idx = 0; idx < 2; ++idx) {
337 if (adev->vce.harvest_config & (1 << idx))
338 continue;
339
340 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(idx));
341
342 if (adev->asic_type >= CHIP_STONEY)
343 WREG32_P(mmVCE_VCPU_CNTL, 0, ~0x200001);
344 else
345 WREG32_FIELD(VCE_VCPU_CNTL, CLK_EN, 0);
346
347 /* hold on ECPU */
348 WREG32_FIELD(VCE_SOFT_RESET, ECPU_SOFT_RESET, 1);
349
350 /* clear VCE STATUS */
351 WREG32(mmVCE_STATUS, 0);
352 }
353
354 WREG32(mmGRBM_GFX_INDEX, mmGRBM_GFX_INDEX_DEFAULT);
355 mutex_unlock(&adev->grbm_idx_mutex);
356
357 return 0;
358 }
359
360 #define ixVCE_HARVEST_FUSE_MACRO__ADDRESS 0xC0014074
361 #define VCE_HARVEST_FUSE_MACRO__SHIFT 27
362 #define VCE_HARVEST_FUSE_MACRO__MASK 0x18000000
363
vce_v3_0_get_harvest_config(struct amdgpu_device * adev)364 static unsigned vce_v3_0_get_harvest_config(struct amdgpu_device *adev)
365 {
366 u32 tmp;
367
368 if ((adev->asic_type == CHIP_FIJI) ||
369 (adev->asic_type == CHIP_STONEY))
370 return AMDGPU_VCE_HARVEST_VCE1;
371
372 if (adev->flags & AMD_IS_APU)
373 tmp = (RREG32_SMC(ixVCE_HARVEST_FUSE_MACRO__ADDRESS) &
374 VCE_HARVEST_FUSE_MACRO__MASK) >>
375 VCE_HARVEST_FUSE_MACRO__SHIFT;
376 else
377 tmp = (RREG32_SMC(ixCC_HARVEST_FUSES) &
378 CC_HARVEST_FUSES__VCE_DISABLE_MASK) >>
379 CC_HARVEST_FUSES__VCE_DISABLE__SHIFT;
380
381 switch (tmp) {
382 case 1:
383 return AMDGPU_VCE_HARVEST_VCE0;
384 case 2:
385 return AMDGPU_VCE_HARVEST_VCE1;
386 case 3:
387 return AMDGPU_VCE_HARVEST_VCE0 | AMDGPU_VCE_HARVEST_VCE1;
388 default:
389 if ((adev->asic_type == CHIP_POLARIS10) ||
390 (adev->asic_type == CHIP_POLARIS11) ||
391 (adev->asic_type == CHIP_POLARIS12) ||
392 (adev->asic_type == CHIP_VEGAM))
393 return AMDGPU_VCE_HARVEST_VCE1;
394
395 return 0;
396 }
397 }
398
vce_v3_0_early_init(struct amdgpu_ip_block * ip_block)399 static int vce_v3_0_early_init(struct amdgpu_ip_block *ip_block)
400 {
401 struct amdgpu_device *adev = ip_block->adev;
402 int r;
403
404 adev->vce.harvest_config = vce_v3_0_get_harvest_config(adev);
405
406 if ((adev->vce.harvest_config &
407 (AMDGPU_VCE_HARVEST_VCE0 | AMDGPU_VCE_HARVEST_VCE1)) ==
408 (AMDGPU_VCE_HARVEST_VCE0 | AMDGPU_VCE_HARVEST_VCE1))
409 return -ENOENT;
410
411 r = amdgpu_vce_early_init(adev);
412 if (r)
413 return r;
414
415 adev->vce.num_rings = 3;
416
417 vce_v3_0_set_ring_funcs(adev);
418 vce_v3_0_set_irq_funcs(adev);
419
420 return 0;
421 }
422
vce_v3_0_sw_init(struct amdgpu_ip_block * ip_block)423 static int vce_v3_0_sw_init(struct amdgpu_ip_block *ip_block)
424 {
425 struct amdgpu_device *adev = ip_block->adev;
426 struct amdgpu_ring *ring;
427 int r, i;
428
429 /* VCE */
430 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_VCE_TRAP, &adev->vce.irq);
431 if (r)
432 return r;
433
434 r = amdgpu_vce_sw_init(adev, VCE_V3_0_FW_SIZE +
435 (VCE_V3_0_STACK_SIZE + VCE_V3_0_DATA_SIZE) * 2);
436 if (r)
437 return r;
438
439 /* 52.8.3 required for 3 ring support */
440 if (adev->vce.fw_version < FW_52_8_3)
441 adev->vce.num_rings = 2;
442
443 r = amdgpu_vce_resume(adev);
444 if (r)
445 return r;
446
447 for (i = 0; i < adev->vce.num_rings; i++) {
448 enum amdgpu_ring_priority_level hw_prio = amdgpu_vce_get_ring_prio(i);
449
450 ring = &adev->vce.ring[i];
451 sprintf(ring->name, "vce%d", i);
452 r = amdgpu_ring_init(adev, ring, 512, &adev->vce.irq, 0,
453 hw_prio, NULL);
454 if (r)
455 return r;
456 }
457
458 return r;
459 }
460
vce_v3_0_sw_fini(struct amdgpu_ip_block * ip_block)461 static int vce_v3_0_sw_fini(struct amdgpu_ip_block *ip_block)
462 {
463 int r;
464 struct amdgpu_device *adev = ip_block->adev;
465
466 r = amdgpu_vce_suspend(adev);
467 if (r)
468 return r;
469
470 return amdgpu_vce_sw_fini(adev);
471 }
472
vce_v3_0_hw_init(struct amdgpu_ip_block * ip_block)473 static int vce_v3_0_hw_init(struct amdgpu_ip_block *ip_block)
474 {
475 int r, i;
476 struct amdgpu_device *adev = ip_block->adev;
477
478 vce_v3_0_override_vce_clock_gating(adev, true);
479
480 amdgpu_asic_set_vce_clocks(adev, 10000, 10000);
481
482 for (i = 0; i < adev->vce.num_rings; i++) {
483 r = amdgpu_ring_test_helper(&adev->vce.ring[i]);
484 if (r)
485 return r;
486 }
487
488 drm_info(adev_to_drm(adev), "VCE initialized successfully.\n");
489
490 return 0;
491 }
492
vce_v3_0_hw_fini(struct amdgpu_ip_block * ip_block)493 static int vce_v3_0_hw_fini(struct amdgpu_ip_block *ip_block)
494 {
495 int r;
496 struct amdgpu_device *adev = ip_block->adev;
497
498 cancel_delayed_work_sync(&adev->vce.idle_work);
499
500 r = vce_v3_0_wait_for_idle(ip_block);
501 if (r)
502 return r;
503
504 vce_v3_0_stop(adev);
505 return vce_v3_0_set_clockgating_state(ip_block, AMD_CG_STATE_GATE);
506 }
507
vce_v3_0_suspend(struct amdgpu_ip_block * ip_block)508 static int vce_v3_0_suspend(struct amdgpu_ip_block *ip_block)
509 {
510 int r;
511 struct amdgpu_device *adev = ip_block->adev;
512
513 /*
514 * Proper cleanups before halting the HW engine:
515 * - cancel the delayed idle work
516 * - enable powergating
517 * - enable clockgating
518 * - disable dpm
519 *
520 * TODO: to align with the VCN implementation, move the
521 * jobs for clockgating/powergating/dpm setting to
522 * ->set_powergating_state().
523 */
524 cancel_delayed_work_sync(&adev->vce.idle_work);
525
526 if (adev->pm.dpm_enabled) {
527 amdgpu_dpm_enable_vce(adev, false);
528 } else {
529 amdgpu_asic_set_vce_clocks(adev, 0, 0);
530 amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
531 AMD_PG_STATE_GATE);
532 amdgpu_device_ip_set_clockgating_state(adev, AMD_IP_BLOCK_TYPE_VCE,
533 AMD_CG_STATE_GATE);
534 }
535
536 r = vce_v3_0_hw_fini(ip_block);
537 if (r)
538 return r;
539
540 return amdgpu_vce_suspend(adev);
541 }
542
vce_v3_0_resume(struct amdgpu_ip_block * ip_block)543 static int vce_v3_0_resume(struct amdgpu_ip_block *ip_block)
544 {
545 int r;
546
547 r = amdgpu_vce_resume(ip_block->adev);
548 if (r)
549 return r;
550
551 return vce_v3_0_hw_init(ip_block);
552 }
553
vce_v3_0_mc_resume(struct amdgpu_device * adev,int idx)554 static void vce_v3_0_mc_resume(struct amdgpu_device *adev, int idx)
555 {
556 uint32_t offset, size;
557
558 WREG32_P(mmVCE_CLOCK_GATING_A, 0, ~(1 << 16));
559 WREG32_P(mmVCE_UENC_CLOCK_GATING, 0x1FF000, ~0xFF9FF000);
560 WREG32_P(mmVCE_UENC_REG_CLOCK_GATING, 0x3F, ~0x3F);
561 WREG32(mmVCE_CLOCK_GATING_B, 0x1FF);
562
563 WREG32(mmVCE_LMI_CTRL, 0x00398000);
564 WREG32_P(mmVCE_LMI_CACHE_CTRL, 0x0, ~0x1);
565 WREG32(mmVCE_LMI_SWAP_CNTL, 0);
566 WREG32(mmVCE_LMI_SWAP_CNTL1, 0);
567 WREG32(mmVCE_LMI_VM_CTRL, 0);
568 WREG32_OR(mmVCE_VCPU_CNTL, 0x00100000);
569
570 if (adev->asic_type >= CHIP_STONEY) {
571 WREG32(mmVCE_LMI_VCPU_CACHE_40BIT_BAR0, (adev->vce.gpu_addr >> 8));
572 WREG32(mmVCE_LMI_VCPU_CACHE_40BIT_BAR1, (adev->vce.gpu_addr >> 8));
573 WREG32(mmVCE_LMI_VCPU_CACHE_40BIT_BAR2, (adev->vce.gpu_addr >> 8));
574 } else
575 WREG32(mmVCE_LMI_VCPU_CACHE_40BIT_BAR, (adev->vce.gpu_addr >> 8));
576 offset = AMDGPU_VCE_FIRMWARE_OFFSET;
577 size = VCE_V3_0_FW_SIZE - AMDGPU_VCE_FIRMWARE_OFFSET;
578 WREG32(mmVCE_VCPU_CACHE_OFFSET0, offset & 0x7fffffff);
579 WREG32(mmVCE_VCPU_CACHE_SIZE0, size);
580
581 if (idx == 0) {
582 offset += size;
583 size = VCE_V3_0_STACK_SIZE;
584 WREG32(mmVCE_VCPU_CACHE_OFFSET1, offset & 0x7fffffff);
585 WREG32(mmVCE_VCPU_CACHE_SIZE1, size);
586 offset += size;
587 size = VCE_V3_0_DATA_SIZE;
588 WREG32(mmVCE_VCPU_CACHE_OFFSET2, offset & 0x7fffffff);
589 WREG32(mmVCE_VCPU_CACHE_SIZE2, size);
590 } else {
591 offset += size + VCE_V3_0_STACK_SIZE + VCE_V3_0_DATA_SIZE;
592 size = VCE_V3_0_STACK_SIZE;
593 WREG32(mmVCE_VCPU_CACHE_OFFSET1, offset & 0xfffffff);
594 WREG32(mmVCE_VCPU_CACHE_SIZE1, size);
595 offset += size;
596 size = VCE_V3_0_DATA_SIZE;
597 WREG32(mmVCE_VCPU_CACHE_OFFSET2, offset & 0xfffffff);
598 WREG32(mmVCE_VCPU_CACHE_SIZE2, size);
599 }
600
601 WREG32_P(mmVCE_LMI_CTRL2, 0x0, ~0x100);
602 WREG32_FIELD(VCE_SYS_INT_EN, VCE_SYS_INT_TRAP_INTERRUPT_EN, 1);
603 }
604
vce_v3_0_is_idle(struct amdgpu_ip_block * ip_block)605 static bool vce_v3_0_is_idle(struct amdgpu_ip_block *ip_block)
606 {
607 struct amdgpu_device *adev = ip_block->adev;
608 u32 mask = 0;
609
610 mask |= (adev->vce.harvest_config & AMDGPU_VCE_HARVEST_VCE0) ? 0 : SRBM_STATUS2__VCE0_BUSY_MASK;
611 mask |= (adev->vce.harvest_config & AMDGPU_VCE_HARVEST_VCE1) ? 0 : SRBM_STATUS2__VCE1_BUSY_MASK;
612
613 return !(RREG32(mmSRBM_STATUS2) & mask);
614 }
615
vce_v3_0_wait_for_idle(struct amdgpu_ip_block * ip_block)616 static int vce_v3_0_wait_for_idle(struct amdgpu_ip_block *ip_block)
617 {
618 unsigned i;
619 struct amdgpu_device *adev = ip_block->adev;
620
621 for (i = 0; i < adev->usec_timeout; i++)
622 if (vce_v3_0_is_idle(ip_block))
623 return 0;
624
625 return -ETIMEDOUT;
626 }
627
628 #define VCE_STATUS_VCPU_REPORT_AUTO_BUSY_MASK 0x00000008L /* AUTO_BUSY */
629 #define VCE_STATUS_VCPU_REPORT_RB0_BUSY_MASK 0x00000010L /* RB0_BUSY */
630 #define VCE_STATUS_VCPU_REPORT_RB1_BUSY_MASK 0x00000020L /* RB1_BUSY */
631 #define AMDGPU_VCE_STATUS_BUSY_MASK (VCE_STATUS_VCPU_REPORT_AUTO_BUSY_MASK | \
632 VCE_STATUS_VCPU_REPORT_RB0_BUSY_MASK)
633
vce_v3_0_soft_reset(struct amdgpu_ip_block * ip_block)634 static int vce_v3_0_soft_reset(struct amdgpu_ip_block *ip_block)
635 {
636 struct amdgpu_device *adev = ip_block->adev;
637 u32 srbm_soft_reset;
638
639 if (!adev->vce.srbm_soft_reset)
640 return 0;
641 srbm_soft_reset = adev->vce.srbm_soft_reset;
642
643 if (srbm_soft_reset) {
644 u32 tmp;
645
646 tmp = RREG32(mmSRBM_SOFT_RESET);
647 tmp |= srbm_soft_reset;
648 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
649 WREG32(mmSRBM_SOFT_RESET, tmp);
650 tmp = RREG32(mmSRBM_SOFT_RESET);
651
652 udelay(50);
653
654 tmp &= ~srbm_soft_reset;
655 WREG32(mmSRBM_SOFT_RESET, tmp);
656 tmp = RREG32(mmSRBM_SOFT_RESET);
657
658 /* Wait a little for things to settle down */
659 udelay(50);
660 }
661
662 return 0;
663 }
664
vce_v3_0_set_interrupt_state(struct amdgpu_device * adev,struct amdgpu_irq_src * source,unsigned type,enum amdgpu_interrupt_state state)665 static int vce_v3_0_set_interrupt_state(struct amdgpu_device *adev,
666 struct amdgpu_irq_src *source,
667 unsigned type,
668 enum amdgpu_interrupt_state state)
669 {
670 uint32_t val = 0;
671
672 if (state == AMDGPU_IRQ_STATE_ENABLE)
673 val |= VCE_SYS_INT_EN__VCE_SYS_INT_TRAP_INTERRUPT_EN_MASK;
674
675 WREG32_P(mmVCE_SYS_INT_EN, val, ~VCE_SYS_INT_EN__VCE_SYS_INT_TRAP_INTERRUPT_EN_MASK);
676 return 0;
677 }
678
vce_v3_0_process_interrupt(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)679 static int vce_v3_0_process_interrupt(struct amdgpu_device *adev,
680 struct amdgpu_irq_src *source,
681 struct amdgpu_iv_entry *entry)
682 {
683 DRM_DEBUG("IH: VCE\n");
684
685 WREG32_FIELD(VCE_SYS_INT_STATUS, VCE_SYS_INT_TRAP_INTERRUPT_INT, 1);
686
687 switch (entry->src_data[0]) {
688 case 0:
689 case 1:
690 case 2:
691 amdgpu_fence_process(&adev->vce.ring[entry->src_data[0]]);
692 break;
693 default:
694 DRM_ERROR("Unhandled interrupt: %d %d\n",
695 entry->src_id, entry->src_data[0]);
696 break;
697 }
698
699 return 0;
700 }
701
vce_v3_0_set_clockgating_state(struct amdgpu_ip_block * ip_block,enum amd_clockgating_state state)702 static int vce_v3_0_set_clockgating_state(struct amdgpu_ip_block *ip_block,
703 enum amd_clockgating_state state)
704 {
705 struct amdgpu_device *adev = ip_block->adev;
706 bool enable = (state == AMD_CG_STATE_GATE);
707 int i;
708
709 if (!(adev->cg_flags & AMD_CG_SUPPORT_VCE_MGCG))
710 return 0;
711
712 mutex_lock(&adev->grbm_idx_mutex);
713 for (i = 0; i < 2; i++) {
714 /* Program VCE Instance 0 or 1 if not harvested */
715 if (adev->vce.harvest_config & (1 << i))
716 continue;
717
718 WREG32(mmGRBM_GFX_INDEX, GET_VCE_INSTANCE(i));
719
720 if (!enable) {
721 /* initialize VCE_CLOCK_GATING_A: Clock ON/OFF delay */
722 uint32_t data = RREG32(mmVCE_CLOCK_GATING_A);
723 data &= ~(0xf | 0xff0);
724 data |= ((0x0 << 0) | (0x04 << 4));
725 WREG32(mmVCE_CLOCK_GATING_A, data);
726
727 /* initialize VCE_UENC_CLOCK_GATING: Clock ON/OFF delay */
728 data = RREG32(mmVCE_UENC_CLOCK_GATING);
729 data &= ~(0xf | 0xff0);
730 data |= ((0x0 << 0) | (0x04 << 4));
731 WREG32(mmVCE_UENC_CLOCK_GATING, data);
732 }
733
734 vce_v3_0_set_vce_sw_clock_gating(adev, enable);
735 }
736
737 WREG32(mmGRBM_GFX_INDEX, mmGRBM_GFX_INDEX_DEFAULT);
738 mutex_unlock(&adev->grbm_idx_mutex);
739
740 return 0;
741 }
742
vce_v3_0_set_powergating_state(struct amdgpu_ip_block * ip_block,enum amd_powergating_state state)743 static int vce_v3_0_set_powergating_state(struct amdgpu_ip_block *ip_block,
744 enum amd_powergating_state state)
745 {
746 /* This doesn't actually powergate the VCE block.
747 * That's done in the dpm code via the SMC. This
748 * just re-inits the block as necessary. The actual
749 * gating still happens in the dpm code. We should
750 * revisit this when there is a cleaner line between
751 * the smc and the hw blocks
752 */
753 struct amdgpu_device *adev = ip_block->adev;
754 int ret = 0;
755
756 if (state == AMD_PG_STATE_GATE) {
757 ret = vce_v3_0_stop(adev);
758 if (ret)
759 goto out;
760 } else {
761 ret = vce_v3_0_start(adev);
762 if (ret)
763 goto out;
764 }
765
766 out:
767 return ret;
768 }
769
vce_v3_0_get_clockgating_state(struct amdgpu_ip_block * ip_block,u64 * flags)770 static void vce_v3_0_get_clockgating_state(struct amdgpu_ip_block *ip_block, u64 *flags)
771 {
772 struct amdgpu_device *adev = ip_block->adev;
773 int data;
774
775 mutex_lock(&adev->pm.mutex);
776
777 if (adev->flags & AMD_IS_APU)
778 data = RREG32_SMC(ixCURRENT_PG_STATUS_APU);
779 else
780 data = RREG32_SMC(ixCURRENT_PG_STATUS);
781
782 if (data & CURRENT_PG_STATUS__VCE_PG_STATUS_MASK) {
783 drm_info(adev_to_drm(adev), "Cannot get clockgating state when VCE is powergated.\n");
784 goto out;
785 }
786
787 WREG32_FIELD(GRBM_GFX_INDEX, VCE_INSTANCE, 0);
788
789 /* AMD_CG_SUPPORT_VCE_MGCG */
790 data = RREG32(mmVCE_CLOCK_GATING_A);
791 if (data & (0x04 << 4))
792 *flags |= AMD_CG_SUPPORT_VCE_MGCG;
793
794 out:
795 mutex_unlock(&adev->pm.mutex);
796 }
797
vce_v3_0_ring_emit_ib(struct amdgpu_ring * ring,struct amdgpu_job * job,struct amdgpu_ib * ib,uint32_t flags)798 static void vce_v3_0_ring_emit_ib(struct amdgpu_ring *ring,
799 struct amdgpu_job *job,
800 struct amdgpu_ib *ib,
801 uint32_t flags)
802 {
803 unsigned vmid = AMDGPU_JOB_GET_VMID(job);
804
805 amdgpu_ring_write(ring, VCE_CMD_IB_VM);
806 amdgpu_ring_write(ring, vmid);
807 amdgpu_ring_write(ring, lower_32_bits(ib->gpu_addr));
808 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr));
809 amdgpu_ring_write(ring, ib->length_dw);
810 }
811
vce_v3_0_ring_emit_fence(struct amdgpu_ring * ring,u64 addr,u64 seq,unsigned flags)812 static void vce_v3_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
813 u64 seq, unsigned flags)
814 {
815 WARN_ON(flags & AMDGPU_FENCE_FLAG_64BIT);
816
817 amdgpu_ring_write(ring, VCE_CMD_FENCE);
818 amdgpu_ring_write(ring, addr);
819 amdgpu_ring_write(ring, upper_32_bits(addr));
820 amdgpu_ring_write(ring, seq);
821 amdgpu_ring_write(ring, VCE_CMD_TRAP);
822 }
823
vce_v3_0_ring_insert_end(struct amdgpu_ring * ring)824 static void vce_v3_0_ring_insert_end(struct amdgpu_ring *ring)
825 {
826 amdgpu_ring_write(ring, VCE_CMD_END);
827 }
828
vce_v3_0_emit_vm_flush(struct amdgpu_ring * ring,unsigned int vmid,uint64_t pd_addr)829 static void vce_v3_0_emit_vm_flush(struct amdgpu_ring *ring,
830 unsigned int vmid, uint64_t pd_addr)
831 {
832 amdgpu_ring_write(ring, VCE_CMD_UPDATE_PTB);
833 amdgpu_ring_write(ring, vmid);
834 amdgpu_ring_write(ring, pd_addr >> 12);
835
836 amdgpu_ring_write(ring, VCE_CMD_FLUSH_TLB);
837 amdgpu_ring_write(ring, vmid);
838 }
839
vce_v3_0_emit_pipeline_sync(struct amdgpu_ring * ring)840 static void vce_v3_0_emit_pipeline_sync(struct amdgpu_ring *ring)
841 {
842 uint32_t seq = ring->fence_drv.sync_seq;
843 uint64_t addr = ring->fence_drv.gpu_addr;
844
845 amdgpu_ring_write(ring, VCE_CMD_WAIT_GE);
846 amdgpu_ring_write(ring, lower_32_bits(addr));
847 amdgpu_ring_write(ring, upper_32_bits(addr));
848 amdgpu_ring_write(ring, seq);
849 }
850
851 static const struct amd_ip_funcs vce_v3_0_ip_funcs = {
852 .name = "vce_v3_0",
853 .early_init = vce_v3_0_early_init,
854 .sw_init = vce_v3_0_sw_init,
855 .sw_fini = vce_v3_0_sw_fini,
856 .hw_init = vce_v3_0_hw_init,
857 .hw_fini = vce_v3_0_hw_fini,
858 .suspend = vce_v3_0_suspend,
859 .resume = vce_v3_0_resume,
860 .is_idle = vce_v3_0_is_idle,
861 .wait_for_idle = vce_v3_0_wait_for_idle,
862 .soft_reset = vce_v3_0_soft_reset,
863 .set_clockgating_state = vce_v3_0_set_clockgating_state,
864 .set_powergating_state = vce_v3_0_set_powergating_state,
865 .get_clockgating_state = vce_v3_0_get_clockgating_state,
866 };
867
868 static const struct amdgpu_ring_funcs vce_v3_0_ring_phys_funcs = {
869 .type = AMDGPU_RING_TYPE_VCE,
870 .align_mask = 0xf,
871 .nop = VCE_CMD_NO_OP,
872 .support_64bit_ptrs = false,
873 .no_user_fence = true,
874 .get_rptr = vce_v3_0_ring_get_rptr,
875 .get_wptr = vce_v3_0_ring_get_wptr,
876 .set_wptr = vce_v3_0_ring_set_wptr,
877 .parse_cs = amdgpu_vce_ring_parse_cs,
878 .emit_frame_size =
879 4 + /* vce_v3_0_emit_pipeline_sync */
880 6, /* amdgpu_vce_ring_emit_fence x1 no user fence */
881 .emit_ib_size = 4, /* amdgpu_vce_ring_emit_ib */
882 .emit_ib = amdgpu_vce_ring_emit_ib,
883 .emit_fence = amdgpu_vce_ring_emit_fence,
884 .test_ring = amdgpu_vce_ring_test_ring,
885 .test_ib = amdgpu_vce_ring_test_ib,
886 .insert_nop = amdgpu_ring_insert_nop,
887 .pad_ib = amdgpu_ring_generic_pad_ib,
888 .begin_use = amdgpu_vce_ring_begin_use,
889 .end_use = amdgpu_vce_ring_end_use,
890 };
891
892 static const struct amdgpu_ring_funcs vce_v3_0_ring_vm_funcs = {
893 .type = AMDGPU_RING_TYPE_VCE,
894 .align_mask = 0x1f,
895 .nop = VCE_CMD_NO_OP,
896 .support_64bit_ptrs = false,
897 .no_user_fence = true,
898 .get_rptr = vce_v3_0_ring_get_rptr,
899 .get_wptr = vce_v3_0_ring_get_wptr,
900 .set_wptr = vce_v3_0_ring_set_wptr,
901 .patch_cs_in_place = amdgpu_vce_ring_parse_cs_vm,
902 .emit_frame_size =
903 5 + /* vce_v3_0_emit_vm_flush */
904 4 + /* vce_v3_0_emit_pipeline_sync */
905 5 + 5 + /* vce_v3_0_ring_emit_fence x2 vm fence */
906 1, /* vce_v3_0_ring_insert_end */
907 .emit_ib_size = 5, /* vce_v3_0_ring_emit_ib */
908 .emit_ib = vce_v3_0_ring_emit_ib,
909 .emit_vm_flush = vce_v3_0_emit_vm_flush,
910 .emit_pipeline_sync = vce_v3_0_emit_pipeline_sync,
911 .emit_fence = vce_v3_0_ring_emit_fence,
912 .test_ring = amdgpu_vce_ring_test_ring,
913 .test_ib = amdgpu_vce_ring_test_ib,
914 .insert_nop = amdgpu_ring_insert_nop,
915 .insert_end = vce_v3_0_ring_insert_end,
916 .pad_ib = amdgpu_ring_generic_pad_ib,
917 .begin_use = amdgpu_vce_ring_begin_use,
918 .end_use = amdgpu_vce_ring_end_use,
919 };
920
vce_v3_0_set_ring_funcs(struct amdgpu_device * adev)921 static void vce_v3_0_set_ring_funcs(struct amdgpu_device *adev)
922 {
923 int i;
924
925 if (adev->asic_type >= CHIP_STONEY) {
926 for (i = 0; i < adev->vce.num_rings; i++) {
927 adev->vce.ring[i].funcs = &vce_v3_0_ring_vm_funcs;
928 adev->vce.ring[i].me = i;
929 }
930 drm_info(adev_to_drm(adev), "VCE enabled in VM mode\n");
931 } else {
932 for (i = 0; i < adev->vce.num_rings; i++) {
933 adev->vce.ring[i].funcs = &vce_v3_0_ring_phys_funcs;
934 adev->vce.ring[i].me = i;
935 }
936 drm_info(adev_to_drm(adev), "VCE enabled in physical mode\n");
937 }
938 }
939
940 static const struct amdgpu_irq_src_funcs vce_v3_0_irq_funcs = {
941 .set = vce_v3_0_set_interrupt_state,
942 .process = vce_v3_0_process_interrupt,
943 };
944
vce_v3_0_set_irq_funcs(struct amdgpu_device * adev)945 static void vce_v3_0_set_irq_funcs(struct amdgpu_device *adev)
946 {
947 adev->vce.irq.num_types = 1;
948 adev->vce.irq.funcs = &vce_v3_0_irq_funcs;
949 };
950
951 const struct amdgpu_ip_block_version vce_v3_0_ip_block = {
952 .type = AMD_IP_BLOCK_TYPE_VCE,
953 .major = 3,
954 .minor = 0,
955 .rev = 0,
956 .funcs = &vce_v3_0_ip_funcs,
957 };
958
959 const struct amdgpu_ip_block_version vce_v3_1_ip_block = {
960 .type = AMD_IP_BLOCK_TYPE_VCE,
961 .major = 3,
962 .minor = 1,
963 .rev = 0,
964 .funcs = &vce_v3_0_ip_funcs,
965 };
966
967 const struct amdgpu_ip_block_version vce_v3_4_ip_block = {
968 .type = AMD_IP_BLOCK_TYPE_VCE,
969 .major = 3,
970 .minor = 4,
971 .rev = 0,
972 .funcs = &vce_v3_0_ip_funcs,
973 };
974