1 /*
2 * Copyright 2014 Advanced Micro Devices, Inc.
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
13 *
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
21 *
22 */
23 #include <linux/firmware.h>
24 #include <drm/drmP.h>
25 #include <drm/drm_cache.h>
26 #include "amdgpu.h"
27 #include "gmc_v8_0.h"
28 #include "amdgpu_ucode.h"
29 #include "amdgpu_amdkfd.h"
30
31 #include "gmc/gmc_8_1_d.h"
32 #include "gmc/gmc_8_1_sh_mask.h"
33
34 #include "bif/bif_5_0_d.h"
35 #include "bif/bif_5_0_sh_mask.h"
36
37 #include "oss/oss_3_0_d.h"
38 #include "oss/oss_3_0_sh_mask.h"
39
40 #include "dce/dce_10_0_d.h"
41 #include "dce/dce_10_0_sh_mask.h"
42
43 #include "vid.h"
44 #include "vi.h"
45
46 #include "amdgpu_atombios.h"
47
48 #include "ivsrcid/ivsrcid_vislands30.h"
49
50 static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev);
51 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev);
52 static int gmc_v8_0_wait_for_idle(void *handle);
53
54 MODULE_FIRMWARE("amdgpufw_tonga_mc");
55 MODULE_FIRMWARE("amdgpufw_polaris11_mc");
56 MODULE_FIRMWARE("amdgpufw_polaris10_mc");
57 MODULE_FIRMWARE("amdgpufw_polaris12_mc");
58 MODULE_FIRMWARE("amdgpufw_polaris11_k_mc");
59 MODULE_FIRMWARE("amdgpufw_polaris10_k_mc");
60 MODULE_FIRMWARE("amdgpufw_polaris12_k_mc");
61
62 static const u32 golden_settings_tonga_a11[] =
63 {
64 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
65 mmMC_HUB_RDREQ_DMIF_LIMIT, 0x0000007f, 0x00000028,
66 mmMC_HUB_WDP_UMC, 0x00007fb6, 0x00000991,
67 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
68 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
69 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
70 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
71 };
72
73 static const u32 tonga_mgcg_cgcg_init[] =
74 {
75 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
76 };
77
78 static const u32 golden_settings_fiji_a10[] =
79 {
80 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
81 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
82 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
83 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
84 };
85
86 static const u32 fiji_mgcg_cgcg_init[] =
87 {
88 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
89 };
90
91 static const u32 golden_settings_polaris11_a11[] =
92 {
93 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
94 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
95 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
96 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
97 };
98
99 static const u32 golden_settings_polaris10_a11[] =
100 {
101 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
102 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
103 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
104 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
105 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
106 };
107
108 static const u32 cz_mgcg_cgcg_init[] =
109 {
110 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
111 };
112
113 static const u32 stoney_mgcg_cgcg_init[] =
114 {
115 mmATC_MISC_CG, 0xffffffff, 0x000c0200,
116 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
117 };
118
119 static const u32 golden_settings_stoney_common[] =
120 {
121 mmMC_HUB_RDREQ_UVD, MC_HUB_RDREQ_UVD__PRESCALE_MASK, 0x00000004,
122 mmMC_RD_GRP_OTH, MC_RD_GRP_OTH__UVD_MASK, 0x00600000
123 };
124
gmc_v8_0_init_golden_registers(struct amdgpu_device * adev)125 static void gmc_v8_0_init_golden_registers(struct amdgpu_device *adev)
126 {
127 switch (adev->asic_type) {
128 case CHIP_FIJI:
129 amdgpu_device_program_register_sequence(adev,
130 fiji_mgcg_cgcg_init,
131 ARRAY_SIZE(fiji_mgcg_cgcg_init));
132 amdgpu_device_program_register_sequence(adev,
133 golden_settings_fiji_a10,
134 ARRAY_SIZE(golden_settings_fiji_a10));
135 break;
136 case CHIP_TONGA:
137 amdgpu_device_program_register_sequence(adev,
138 tonga_mgcg_cgcg_init,
139 ARRAY_SIZE(tonga_mgcg_cgcg_init));
140 amdgpu_device_program_register_sequence(adev,
141 golden_settings_tonga_a11,
142 ARRAY_SIZE(golden_settings_tonga_a11));
143 break;
144 case CHIP_POLARIS11:
145 case CHIP_POLARIS12:
146 case CHIP_VEGAM:
147 amdgpu_device_program_register_sequence(adev,
148 golden_settings_polaris11_a11,
149 ARRAY_SIZE(golden_settings_polaris11_a11));
150 break;
151 case CHIP_POLARIS10:
152 amdgpu_device_program_register_sequence(adev,
153 golden_settings_polaris10_a11,
154 ARRAY_SIZE(golden_settings_polaris10_a11));
155 break;
156 case CHIP_CARRIZO:
157 amdgpu_device_program_register_sequence(adev,
158 cz_mgcg_cgcg_init,
159 ARRAY_SIZE(cz_mgcg_cgcg_init));
160 break;
161 case CHIP_STONEY:
162 amdgpu_device_program_register_sequence(adev,
163 stoney_mgcg_cgcg_init,
164 ARRAY_SIZE(stoney_mgcg_cgcg_init));
165 amdgpu_device_program_register_sequence(adev,
166 golden_settings_stoney_common,
167 ARRAY_SIZE(golden_settings_stoney_common));
168 break;
169 default:
170 break;
171 }
172 }
173
gmc_v8_0_mc_stop(struct amdgpu_device * adev)174 static void gmc_v8_0_mc_stop(struct amdgpu_device *adev)
175 {
176 u32 blackout;
177
178 gmc_v8_0_wait_for_idle(adev);
179
180 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
181 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) {
182 /* Block CPU access */
183 WREG32(mmBIF_FB_EN, 0);
184 /* blackout the MC */
185 blackout = REG_SET_FIELD(blackout,
186 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 1);
187 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout);
188 }
189 /* wait for the MC to settle */
190 udelay(100);
191 }
192
gmc_v8_0_mc_resume(struct amdgpu_device * adev)193 static void gmc_v8_0_mc_resume(struct amdgpu_device *adev)
194 {
195 u32 tmp;
196
197 /* unblackout the MC */
198 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
199 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
200 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp);
201 /* allow CPU access */
202 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1);
203 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1);
204 WREG32(mmBIF_FB_EN, tmp);
205 }
206
207 /**
208 * gmc_v8_0_init_microcode - load ucode images from disk
209 *
210 * @adev: amdgpu_device pointer
211 *
212 * Use the firmware interface to load the ucode images into
213 * the driver (not loaded into hw).
214 * Returns 0 on success, error on failure.
215 */
gmc_v8_0_init_microcode(struct amdgpu_device * adev)216 static int gmc_v8_0_init_microcode(struct amdgpu_device *adev)
217 {
218 const char *chip_name;
219 char fw_name[30];
220 int err;
221
222 DRM_DEBUG("\n");
223
224 switch (adev->asic_type) {
225 case CHIP_TONGA:
226 chip_name = "tonga";
227 break;
228 case CHIP_POLARIS11:
229 if (((adev->pdev->device == 0x67ef) &&
230 ((adev->pdev->revision == 0xe0) ||
231 (adev->pdev->revision == 0xe5))) ||
232 ((adev->pdev->device == 0x67ff) &&
233 ((adev->pdev->revision == 0xcf) ||
234 (adev->pdev->revision == 0xef) ||
235 (adev->pdev->revision == 0xff))))
236 chip_name = "polaris11_k";
237 else if ((adev->pdev->device == 0x67ef) &&
238 (adev->pdev->revision == 0xe2))
239 chip_name = "polaris11_k";
240 else
241 chip_name = "polaris11";
242 break;
243 case CHIP_POLARIS10:
244 if ((adev->pdev->device == 0x67df) &&
245 ((adev->pdev->revision == 0xe1) ||
246 (adev->pdev->revision == 0xf7)))
247 chip_name = "polaris10_k";
248 else
249 chip_name = "polaris10";
250 break;
251 case CHIP_POLARIS12:
252 if (((adev->pdev->device == 0x6987) &&
253 ((adev->pdev->revision == 0xc0) ||
254 (adev->pdev->revision == 0xc3))) ||
255 ((adev->pdev->device == 0x6981) &&
256 ((adev->pdev->revision == 0x00) ||
257 (adev->pdev->revision == 0x01) ||
258 (adev->pdev->revision == 0x10))))
259 chip_name = "polaris12_k";
260 else
261 chip_name = "polaris12";
262 break;
263 case CHIP_FIJI:
264 case CHIP_CARRIZO:
265 case CHIP_STONEY:
266 case CHIP_VEGAM:
267 return 0;
268 default: BUG();
269 }
270
271 snprintf(fw_name, sizeof(fw_name), "amdgpufw_%s_mc", chip_name);
272 err = request_firmware(&adev->gmc.fw, fw_name, adev->dev);
273 if (err)
274 goto out;
275 err = amdgpu_ucode_validate(adev->gmc.fw);
276
277 out:
278 if (err) {
279 pr_err("mc: Failed to load firmware \"%s\"\n", fw_name);
280 release_firmware(adev->gmc.fw);
281 adev->gmc.fw = NULL;
282 }
283 return err;
284 }
285
286 /**
287 * gmc_v8_0_tonga_mc_load_microcode - load tonga MC ucode into the hw
288 *
289 * @adev: amdgpu_device pointer
290 *
291 * Load the GDDR MC ucode into the hw (CIK).
292 * Returns 0 on success, error on failure.
293 */
gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device * adev)294 static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
295 {
296 const struct mc_firmware_header_v1_0 *hdr;
297 const __le32 *fw_data = NULL;
298 const __le32 *io_mc_regs = NULL;
299 u32 running;
300 int i, ucode_size, regs_size;
301
302 /* Skip MC ucode loading on SR-IOV capable boards.
303 * vbios does this for us in asic_init in that case.
304 * Skip MC ucode loading on VF, because hypervisor will do that
305 * for this adaptor.
306 */
307 if (amdgpu_sriov_bios(adev))
308 return 0;
309
310 if (!adev->gmc.fw)
311 return -EINVAL;
312
313 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
314 amdgpu_ucode_print_mc_hdr(&hdr->header);
315
316 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
317 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
318 io_mc_regs = (const __le32 *)
319 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
320 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
321 fw_data = (const __le32 *)
322 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
323
324 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN);
325
326 if (running == 0) {
327 /* reset the engine and set to writable */
328 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
329 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
330
331 /* load mc io regs */
332 for (i = 0; i < regs_size; i++) {
333 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
334 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
335 }
336 /* load the MC ucode */
337 for (i = 0; i < ucode_size; i++)
338 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
339
340 /* put the engine back into the active state */
341 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
342 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
343 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
344
345 /* wait for training to complete */
346 for (i = 0; i < adev->usec_timeout; i++) {
347 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
348 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0))
349 break;
350 udelay(1);
351 }
352 for (i = 0; i < adev->usec_timeout; i++) {
353 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
354 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1))
355 break;
356 udelay(1);
357 }
358 }
359
360 return 0;
361 }
362
gmc_v8_0_polaris_mc_load_microcode(struct amdgpu_device * adev)363 static int gmc_v8_0_polaris_mc_load_microcode(struct amdgpu_device *adev)
364 {
365 const struct mc_firmware_header_v1_0 *hdr;
366 const __le32 *fw_data = NULL;
367 const __le32 *io_mc_regs = NULL;
368 u32 data;
369 int i, ucode_size, regs_size;
370
371 /* Skip MC ucode loading on SR-IOV capable boards.
372 * vbios does this for us in asic_init in that case.
373 * Skip MC ucode loading on VF, because hypervisor will do that
374 * for this adaptor.
375 */
376 if (amdgpu_sriov_bios(adev))
377 return 0;
378
379 if (!adev->gmc.fw)
380 return -EINVAL;
381
382 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
383 amdgpu_ucode_print_mc_hdr(&hdr->header);
384
385 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
386 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
387 io_mc_regs = (const __le32 *)
388 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
389 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
390 fw_data = (const __le32 *)
391 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
392
393 data = RREG32(mmMC_SEQ_MISC0);
394 data &= ~(0x40);
395 WREG32(mmMC_SEQ_MISC0, data);
396
397 /* load mc io regs */
398 for (i = 0; i < regs_size; i++) {
399 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
400 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
401 }
402
403 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
404 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
405
406 /* load the MC ucode */
407 for (i = 0; i < ucode_size; i++)
408 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
409
410 /* put the engine back into the active state */
411 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
412 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
413 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
414
415 /* wait for training to complete */
416 for (i = 0; i < adev->usec_timeout; i++) {
417 data = RREG32(mmMC_SEQ_MISC0);
418 if (data & 0x80)
419 break;
420 udelay(1);
421 }
422
423 return 0;
424 }
425
gmc_v8_0_vram_gtt_location(struct amdgpu_device * adev,struct amdgpu_gmc * mc)426 static void gmc_v8_0_vram_gtt_location(struct amdgpu_device *adev,
427 struct amdgpu_gmc *mc)
428 {
429 u64 base = 0;
430
431 if (!amdgpu_sriov_vf(adev))
432 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF;
433 base <<= 24;
434
435 amdgpu_device_vram_location(adev, &adev->gmc, base);
436 amdgpu_device_gart_location(adev, mc);
437 }
438
439 /**
440 * gmc_v8_0_mc_program - program the GPU memory controller
441 *
442 * @adev: amdgpu_device pointer
443 *
444 * Set the location of vram, gart, and AGP in the GPU's
445 * physical address space (CIK).
446 */
gmc_v8_0_mc_program(struct amdgpu_device * adev)447 static void gmc_v8_0_mc_program(struct amdgpu_device *adev)
448 {
449 u32 tmp;
450 int i, j;
451
452 /* Initialize HDP */
453 for (i = 0, j = 0; i < 32; i++, j += 0x6) {
454 WREG32((0xb05 + j), 0x00000000);
455 WREG32((0xb06 + j), 0x00000000);
456 WREG32((0xb07 + j), 0x00000000);
457 WREG32((0xb08 + j), 0x00000000);
458 WREG32((0xb09 + j), 0x00000000);
459 }
460 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0);
461
462 if (gmc_v8_0_wait_for_idle((void *)adev)) {
463 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
464 }
465 if (adev->mode_info.num_crtc) {
466 /* Lockout access through VGA aperture*/
467 tmp = RREG32(mmVGA_HDP_CONTROL);
468 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
469 WREG32(mmVGA_HDP_CONTROL, tmp);
470
471 /* disable VGA render */
472 tmp = RREG32(mmVGA_RENDER_CONTROL);
473 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
474 WREG32(mmVGA_RENDER_CONTROL, tmp);
475 }
476 /* Update configuration */
477 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR,
478 adev->gmc.vram_start >> 12);
479 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR,
480 adev->gmc.vram_end >> 12);
481 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
482 adev->vram_scratch.gpu_addr >> 12);
483
484 if (amdgpu_sriov_vf(adev)) {
485 tmp = ((adev->gmc.vram_end >> 24) & 0xFFFF) << 16;
486 tmp |= ((adev->gmc.vram_start >> 24) & 0xFFFF);
487 WREG32(mmMC_VM_FB_LOCATION, tmp);
488 /* XXX double check these! */
489 WREG32(mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
490 WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
491 WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF);
492 }
493
494 WREG32(mmMC_VM_AGP_BASE, 0);
495 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF);
496 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF);
497 if (gmc_v8_0_wait_for_idle((void *)adev)) {
498 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
499 }
500
501 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
502
503 tmp = RREG32(mmHDP_MISC_CNTL);
504 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0);
505 WREG32(mmHDP_MISC_CNTL, tmp);
506
507 tmp = RREG32(mmHDP_HOST_PATH_CNTL);
508 WREG32(mmHDP_HOST_PATH_CNTL, tmp);
509 }
510
511 /**
512 * gmc_v8_0_mc_init - initialize the memory controller driver params
513 *
514 * @adev: amdgpu_device pointer
515 *
516 * Look up the amount of vram, vram width, and decide how to place
517 * vram and gart within the GPU's physical address space (CIK).
518 * Returns 0 for success.
519 */
gmc_v8_0_mc_init(struct amdgpu_device * adev)520 static int gmc_v8_0_mc_init(struct amdgpu_device *adev)
521 {
522 int r;
523
524 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev);
525 if (!adev->gmc.vram_width) {
526 u32 tmp;
527 int chansize, numchan;
528
529 /* Get VRAM informations */
530 tmp = RREG32(mmMC_ARB_RAMCFG);
531 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) {
532 chansize = 64;
533 } else {
534 chansize = 32;
535 }
536 tmp = RREG32(mmMC_SHARED_CHMAP);
537 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) {
538 case 0:
539 default:
540 numchan = 1;
541 break;
542 case 1:
543 numchan = 2;
544 break;
545 case 2:
546 numchan = 4;
547 break;
548 case 3:
549 numchan = 8;
550 break;
551 case 4:
552 numchan = 3;
553 break;
554 case 5:
555 numchan = 6;
556 break;
557 case 6:
558 numchan = 10;
559 break;
560 case 7:
561 numchan = 12;
562 break;
563 case 8:
564 numchan = 16;
565 break;
566 }
567 adev->gmc.vram_width = numchan * chansize;
568 }
569 /* size in MB on si */
570 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
571 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
572
573 if (!(adev->flags & AMD_IS_APU)) {
574 r = amdgpu_device_resize_fb_bar(adev);
575 if (r)
576 return r;
577 }
578 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
579 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
580
581 #ifdef CONFIG_X86_64
582 if (adev->flags & AMD_IS_APU) {
583 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22;
584 adev->gmc.aper_size = adev->gmc.real_vram_size;
585 }
586 #endif
587
588 /* In case the PCI BAR is larger than the actual amount of vram */
589 adev->gmc.visible_vram_size = adev->gmc.aper_size;
590 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
591 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
592
593 /* set the gart size */
594 if (amdgpu_gart_size == -1) {
595 switch (adev->asic_type) {
596 case CHIP_POLARIS10: /* all engines support GPUVM */
597 case CHIP_POLARIS11: /* all engines support GPUVM */
598 case CHIP_POLARIS12: /* all engines support GPUVM */
599 case CHIP_VEGAM: /* all engines support GPUVM */
600 default:
601 adev->gmc.gart_size = 256ULL << 20;
602 break;
603 case CHIP_TONGA: /* UVD, VCE do not support GPUVM */
604 case CHIP_FIJI: /* UVD, VCE do not support GPUVM */
605 case CHIP_CARRIZO: /* UVD, VCE do not support GPUVM, DCE SG support */
606 case CHIP_STONEY: /* UVD does not support GPUVM, DCE SG support */
607 adev->gmc.gart_size = 1024ULL << 20;
608 break;
609 }
610 } else {
611 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
612 }
613
614 gmc_v8_0_vram_gtt_location(adev, &adev->gmc);
615
616 return 0;
617 }
618
619 /*
620 * GART
621 * VMID 0 is the physical GPU addresses as used by the kernel.
622 * VMIDs 1-15 are used for userspace clients and are handled
623 * by the amdgpu vm/hsa code.
624 */
625
626 /**
627 * gmc_v8_0_flush_gpu_tlb - gart tlb flush callback
628 *
629 * @adev: amdgpu_device pointer
630 * @vmid: vm instance to flush
631 *
632 * Flush the TLB for the requested page table (CIK).
633 */
gmc_v8_0_flush_gpu_tlb(struct amdgpu_device * adev,uint32_t vmid)634 static void gmc_v8_0_flush_gpu_tlb(struct amdgpu_device *adev,
635 uint32_t vmid)
636 {
637 /* bits 0-15 are the VM contexts0-15 */
638 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
639 }
640
gmc_v8_0_emit_flush_gpu_tlb(struct amdgpu_ring * ring,unsigned vmid,uint64_t pd_addr)641 static uint64_t gmc_v8_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
642 unsigned vmid, uint64_t pd_addr)
643 {
644 uint32_t reg;
645
646 if (vmid < 8)
647 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid;
648 else
649 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8;
650 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12);
651
652 /* bits 0-15 are the VM contexts0-15 */
653 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid);
654
655 return pd_addr;
656 }
657
gmc_v8_0_emit_pasid_mapping(struct amdgpu_ring * ring,unsigned vmid,unsigned pasid)658 static void gmc_v8_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
659 unsigned pasid)
660 {
661 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid);
662 }
663
664 /**
665 * gmc_v8_0_set_pte_pde - update the page tables using MMIO
666 *
667 * @adev: amdgpu_device pointer
668 * @cpu_pt_addr: cpu address of the page table
669 * @gpu_page_idx: entry in the page table to update
670 * @addr: dst addr to write into pte/pde
671 * @flags: access flags
672 *
673 * Update the page tables using the CPU.
674 */
gmc_v8_0_set_pte_pde(struct amdgpu_device * adev,void * cpu_pt_addr,uint32_t gpu_page_idx,uint64_t addr,uint64_t flags)675 static int gmc_v8_0_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
676 uint32_t gpu_page_idx, uint64_t addr,
677 uint64_t flags)
678 {
679 void __iomem *ptr = (void *)cpu_pt_addr;
680 uint64_t value;
681
682 /*
683 * PTE format on VI:
684 * 63:40 reserved
685 * 39:12 4k physical page base address
686 * 11:7 fragment
687 * 6 write
688 * 5 read
689 * 4 exe
690 * 3 reserved
691 * 2 snooped
692 * 1 system
693 * 0 valid
694 *
695 * PDE format on VI:
696 * 63:59 block fragment size
697 * 58:40 reserved
698 * 39:1 physical base address of PTE
699 * bits 5:1 must be 0.
700 * 0 valid
701 */
702 value = addr & 0x000000FFFFFFF000ULL;
703 value |= flags;
704 writeq(value, ptr + (gpu_page_idx * 8));
705
706 return 0;
707 }
708
gmc_v8_0_get_vm_pte_flags(struct amdgpu_device * adev,uint32_t flags)709 static uint64_t gmc_v8_0_get_vm_pte_flags(struct amdgpu_device *adev,
710 uint32_t flags)
711 {
712 uint64_t pte_flag = 0;
713
714 if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
715 pte_flag |= AMDGPU_PTE_EXECUTABLE;
716 if (flags & AMDGPU_VM_PAGE_READABLE)
717 pte_flag |= AMDGPU_PTE_READABLE;
718 if (flags & AMDGPU_VM_PAGE_WRITEABLE)
719 pte_flag |= AMDGPU_PTE_WRITEABLE;
720 if (flags & AMDGPU_VM_PAGE_PRT)
721 pte_flag |= AMDGPU_PTE_PRT;
722
723 return pte_flag;
724 }
725
gmc_v8_0_get_vm_pde(struct amdgpu_device * adev,int level,uint64_t * addr,uint64_t * flags)726 static void gmc_v8_0_get_vm_pde(struct amdgpu_device *adev, int level,
727 uint64_t *addr, uint64_t *flags)
728 {
729 BUG_ON(*addr & 0xFFFFFF0000000FFFULL);
730 }
731
732 /**
733 * gmc_v8_0_set_fault_enable_default - update VM fault handling
734 *
735 * @adev: amdgpu_device pointer
736 * @value: true redirects VM faults to the default page
737 */
gmc_v8_0_set_fault_enable_default(struct amdgpu_device * adev,bool value)738 static void gmc_v8_0_set_fault_enable_default(struct amdgpu_device *adev,
739 bool value)
740 {
741 u32 tmp;
742
743 tmp = RREG32(mmVM_CONTEXT1_CNTL);
744 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
745 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
746 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
747 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
748 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
749 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
750 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
751 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
752 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
753 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
754 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
755 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
756 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
757 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
758 WREG32(mmVM_CONTEXT1_CNTL, tmp);
759 }
760
761 /**
762 * gmc_v8_0_set_prt - set PRT VM fault
763 *
764 * @adev: amdgpu_device pointer
765 * @enable: enable/disable VM fault handling for PRT
766 */
gmc_v8_0_set_prt(struct amdgpu_device * adev,bool enable)767 static void gmc_v8_0_set_prt(struct amdgpu_device *adev, bool enable)
768 {
769 u32 tmp;
770
771 if (enable && !adev->gmc.prt_warning) {
772 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n");
773 adev->gmc.prt_warning = true;
774 }
775
776 tmp = RREG32(mmVM_PRT_CNTL);
777 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
778 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
779 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
780 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
781 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
782 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
783 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
784 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
785 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
786 L2_CACHE_STORE_INVALID_ENTRIES, enable);
787 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
788 L1_TLB_STORE_INVALID_ENTRIES, enable);
789 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
790 MASK_PDE0_FAULT, enable);
791 WREG32(mmVM_PRT_CNTL, tmp);
792
793 if (enable) {
794 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT;
795 uint32_t high = adev->vm_manager.max_pfn -
796 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT);
797
798 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low);
799 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low);
800 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low);
801 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low);
802 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high);
803 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high);
804 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high);
805 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high);
806 } else {
807 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff);
808 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff);
809 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff);
810 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff);
811 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0);
812 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0);
813 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0);
814 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0);
815 }
816 }
817
818 /**
819 * gmc_v8_0_gart_enable - gart enable
820 *
821 * @adev: amdgpu_device pointer
822 *
823 * This sets up the TLBs, programs the page tables for VMID0,
824 * sets up the hw for VMIDs 1-15 which are allocated on
825 * demand, and sets up the global locations for the LDS, GDS,
826 * and GPUVM for FSA64 clients (CIK).
827 * Returns 0 for success, errors for failure.
828 */
gmc_v8_0_gart_enable(struct amdgpu_device * adev)829 static int gmc_v8_0_gart_enable(struct amdgpu_device *adev)
830 {
831 int r, i;
832 u32 tmp, field;
833
834 if (adev->gart.robj == NULL) {
835 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
836 return -EINVAL;
837 }
838 r = amdgpu_gart_table_vram_pin(adev);
839 if (r)
840 return r;
841 /* Setup TLB control */
842 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
843 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
844 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
845 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
846 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
847 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
848 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
849 /* Setup L2 cache */
850 tmp = RREG32(mmVM_L2_CNTL);
851 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
852 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
853 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
854 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
855 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
856 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
857 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1);
858 WREG32(mmVM_L2_CNTL, tmp);
859 tmp = RREG32(mmVM_L2_CNTL2);
860 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
861 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
862 WREG32(mmVM_L2_CNTL2, tmp);
863
864 field = adev->vm_manager.fragment_size;
865 tmp = RREG32(mmVM_L2_CNTL3);
866 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
867 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field);
868 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field);
869 WREG32(mmVM_L2_CNTL3, tmp);
870 /* XXX: set to enable PTE/PDE in system memory */
871 tmp = RREG32(mmVM_L2_CNTL4);
872 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_PHYSICAL, 0);
873 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SHARED, 0);
874 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SNOOP, 0);
875 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_PHYSICAL, 0);
876 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SHARED, 0);
877 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SNOOP, 0);
878 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_PHYSICAL, 0);
879 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SHARED, 0);
880 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SNOOP, 0);
881 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_PHYSICAL, 0);
882 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SHARED, 0);
883 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SNOOP, 0);
884 WREG32(mmVM_L2_CNTL4, tmp);
885 /* setup context0 */
886 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12);
887 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12);
888 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, adev->gart.table_addr >> 12);
889 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
890 (u32)(adev->dummy_page_addr >> 12));
891 WREG32(mmVM_CONTEXT0_CNTL2, 0);
892 tmp = RREG32(mmVM_CONTEXT0_CNTL);
893 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
894 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
895 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
896 WREG32(mmVM_CONTEXT0_CNTL, tmp);
897
898 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR, 0);
899 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR, 0);
900 WREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET, 0);
901
902 /* empty context1-15 */
903 /* FIXME start with 4G, once using 2 level pt switch to full
904 * vm size space
905 */
906 /* set vm size, must be a multiple of 4 */
907 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
908 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
909 for (i = 1; i < 16; i++) {
910 if (i < 8)
911 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
912 adev->gart.table_addr >> 12);
913 else
914 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
915 adev->gart.table_addr >> 12);
916 }
917
918 /* enable context1-15 */
919 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
920 (u32)(adev->dummy_page_addr >> 12));
921 WREG32(mmVM_CONTEXT1_CNTL2, 4);
922 tmp = RREG32(mmVM_CONTEXT1_CNTL);
923 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
924 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
925 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
926 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
927 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
928 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
929 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
930 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
931 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
932 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
933 adev->vm_manager.block_size - 9);
934 WREG32(mmVM_CONTEXT1_CNTL, tmp);
935 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
936 gmc_v8_0_set_fault_enable_default(adev, false);
937 else
938 gmc_v8_0_set_fault_enable_default(adev, true);
939
940 gmc_v8_0_flush_gpu_tlb(adev, 0);
941 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
942 (unsigned)(adev->gmc.gart_size >> 20),
943 (unsigned long long)adev->gart.table_addr);
944 adev->gart.ready = true;
945 return 0;
946 }
947
gmc_v8_0_gart_init(struct amdgpu_device * adev)948 static int gmc_v8_0_gart_init(struct amdgpu_device *adev)
949 {
950 int r;
951
952 if (adev->gart.robj) {
953 WARN(1, "R600 PCIE GART already initialized\n");
954 return 0;
955 }
956 /* Initialize common gart structure */
957 r = amdgpu_gart_init(adev);
958 if (r)
959 return r;
960 adev->gart.table_size = adev->gart.num_gpu_pages * 8;
961 adev->gart.gart_pte_flags = AMDGPU_PTE_EXECUTABLE;
962 return amdgpu_gart_table_vram_alloc(adev);
963 }
964
965 /**
966 * gmc_v8_0_gart_disable - gart disable
967 *
968 * @adev: amdgpu_device pointer
969 *
970 * This disables all VM page table (CIK).
971 */
gmc_v8_0_gart_disable(struct amdgpu_device * adev)972 static void gmc_v8_0_gart_disable(struct amdgpu_device *adev)
973 {
974 u32 tmp;
975
976 /* Disable all tables */
977 WREG32(mmVM_CONTEXT0_CNTL, 0);
978 WREG32(mmVM_CONTEXT1_CNTL, 0);
979 /* Setup TLB control */
980 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
981 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
982 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
983 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
984 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
985 /* Setup L2 cache */
986 tmp = RREG32(mmVM_L2_CNTL);
987 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
988 WREG32(mmVM_L2_CNTL, tmp);
989 WREG32(mmVM_L2_CNTL2, 0);
990 amdgpu_gart_table_vram_unpin(adev);
991 }
992
993 /**
994 * gmc_v8_0_vm_decode_fault - print human readable fault info
995 *
996 * @adev: amdgpu_device pointer
997 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
998 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
999 *
1000 * Print human readable fault information (CIK).
1001 */
gmc_v8_0_vm_decode_fault(struct amdgpu_device * adev,u32 status,u32 addr,u32 mc_client,unsigned pasid)1002 static void gmc_v8_0_vm_decode_fault(struct amdgpu_device *adev, u32 status,
1003 u32 addr, u32 mc_client, unsigned pasid)
1004 {
1005 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
1006 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1007 PROTECTIONS);
1008 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
1009 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
1010 u32 mc_id;
1011
1012 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1013 MEMORY_CLIENT_ID);
1014
1015 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
1016 protections, vmid, pasid, addr,
1017 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1018 MEMORY_CLIENT_RW) ?
1019 "write" : "read", block, mc_client, mc_id);
1020 }
1021
gmc_v8_0_convert_vram_type(int mc_seq_vram_type)1022 static int gmc_v8_0_convert_vram_type(int mc_seq_vram_type)
1023 {
1024 switch (mc_seq_vram_type) {
1025 case MC_SEQ_MISC0__MT__GDDR1:
1026 return AMDGPU_VRAM_TYPE_GDDR1;
1027 case MC_SEQ_MISC0__MT__DDR2:
1028 return AMDGPU_VRAM_TYPE_DDR2;
1029 case MC_SEQ_MISC0__MT__GDDR3:
1030 return AMDGPU_VRAM_TYPE_GDDR3;
1031 case MC_SEQ_MISC0__MT__GDDR4:
1032 return AMDGPU_VRAM_TYPE_GDDR4;
1033 case MC_SEQ_MISC0__MT__GDDR5:
1034 return AMDGPU_VRAM_TYPE_GDDR5;
1035 case MC_SEQ_MISC0__MT__HBM:
1036 return AMDGPU_VRAM_TYPE_HBM;
1037 case MC_SEQ_MISC0__MT__DDR3:
1038 return AMDGPU_VRAM_TYPE_DDR3;
1039 default:
1040 return AMDGPU_VRAM_TYPE_UNKNOWN;
1041 }
1042 }
1043
gmc_v8_0_early_init(void * handle)1044 static int gmc_v8_0_early_init(void *handle)
1045 {
1046 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1047
1048 gmc_v8_0_set_gmc_funcs(adev);
1049 gmc_v8_0_set_irq_funcs(adev);
1050
1051 adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1052 adev->gmc.shared_aperture_end =
1053 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1054 adev->gmc.private_aperture_start =
1055 adev->gmc.shared_aperture_end + 1;
1056 adev->gmc.private_aperture_end =
1057 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1058
1059 return 0;
1060 }
1061
gmc_v8_0_late_init(void * handle)1062 static int gmc_v8_0_late_init(void *handle)
1063 {
1064 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1065
1066 amdgpu_bo_late_init(adev);
1067
1068 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS)
1069 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1070 else
1071 return 0;
1072 }
1073
gmc_v8_0_get_vbios_fb_size(struct amdgpu_device * adev)1074 static unsigned gmc_v8_0_get_vbios_fb_size(struct amdgpu_device *adev)
1075 {
1076 u32 d1vga_control = RREG32(mmD1VGA_CONTROL);
1077 unsigned size;
1078
1079 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1080 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
1081 } else {
1082 u32 viewport = RREG32(mmVIEWPORT_SIZE);
1083 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1084 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1085 4);
1086 }
1087 /* return 0 if the pre-OS buffer uses up most of vram */
1088 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1089 return 0;
1090 return size;
1091 }
1092
1093 #define mmMC_SEQ_MISC0_FIJI 0xA71
1094
gmc_v8_0_sw_init(void * handle)1095 static int gmc_v8_0_sw_init(void *handle)
1096 {
1097 int r;
1098 int dma_bits;
1099 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1100
1101 if (adev->flags & AMD_IS_APU) {
1102 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
1103 } else {
1104 u32 tmp;
1105
1106 if ((adev->asic_type == CHIP_FIJI) ||
1107 (adev->asic_type == CHIP_VEGAM))
1108 tmp = RREG32(mmMC_SEQ_MISC0_FIJI);
1109 else
1110 tmp = RREG32(mmMC_SEQ_MISC0);
1111 tmp &= MC_SEQ_MISC0__MT__MASK;
1112 adev->gmc.vram_type = gmc_v8_0_convert_vram_type(tmp);
1113 }
1114
1115 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault);
1116 if (r)
1117 return r;
1118
1119 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault);
1120 if (r)
1121 return r;
1122
1123 /* Adjust VM size here.
1124 * Currently set to 4GB ((1 << 20) 4k pages).
1125 * Max GPUVM size for cayman and SI is 40 bits.
1126 */
1127 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40);
1128
1129 /* Set the internal MC address mask
1130 * This is the max address of the GPU's
1131 * internal address space.
1132 */
1133 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
1134
1135 /* set DMA mask + need_dma32 flags.
1136 * PCIE - can handle 40-bits.
1137 * IGP - can handle 40-bits
1138 * PCI - dma32 for legacy pci gart, 40 bits on newer asics
1139 */
1140 adev->need_dma32 = false;
1141 dma_bits = adev->need_dma32 ? 32 : 40;
1142 r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
1143 if (r) {
1144 adev->need_dma32 = true;
1145 dma_bits = 32;
1146 pr_warn("amdgpu: No suitable DMA available\n");
1147 }
1148 r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
1149 if (r) {
1150 pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32));
1151 pr_warn("amdgpu: No coherent DMA available\n");
1152 }
1153 #if 0
1154 adev->need_swiotlb = drm_get_max_iomem() > ((u64)1 << dma_bits);
1155 #endif
1156 /* XXX DragonFly: FreeBSD implementation returns false
1157 * drm-kmod excerpt:
1158 * Only used in combination with CONFIG_SWIOTLB in v4.17
1159 * BSDFIXME: Let's say we can dma all physical memory...
1160 */
1161 adev->need_swiotlb = false;
1162
1163 r = gmc_v8_0_init_microcode(adev);
1164 if (r) {
1165 DRM_ERROR("Failed to load mc firmware!\n");
1166 return r;
1167 }
1168
1169 r = gmc_v8_0_mc_init(adev);
1170 if (r)
1171 return r;
1172
1173 adev->gmc.stolen_size = gmc_v8_0_get_vbios_fb_size(adev);
1174
1175 /* Memory manager */
1176 r = amdgpu_bo_init(adev);
1177 if (r)
1178 return r;
1179
1180 r = gmc_v8_0_gart_init(adev);
1181 if (r)
1182 return r;
1183
1184 /*
1185 * number of VMs
1186 * VMID 0 is reserved for System
1187 * amdgpu graphics/compute will use VMIDs 1-7
1188 * amdkfd will use VMIDs 8-15
1189 */
1190 adev->vm_manager.id_mgr[0].num_ids = AMDGPU_NUM_OF_VMIDS;
1191 amdgpu_vm_manager_init(adev);
1192
1193 /* base offset of vram pages */
1194 if (adev->flags & AMD_IS_APU) {
1195 u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
1196
1197 tmp <<= 22;
1198 adev->vm_manager.vram_base_offset = tmp;
1199 } else {
1200 adev->vm_manager.vram_base_offset = 0;
1201 }
1202
1203 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info),
1204 M_DRM,
1205 GFP_KERNEL);
1206 if (!adev->gmc.vm_fault_info)
1207 return -ENOMEM;
1208 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1209
1210 return 0;
1211 }
1212
gmc_v8_0_sw_fini(void * handle)1213 static int gmc_v8_0_sw_fini(void *handle)
1214 {
1215 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1216
1217 amdgpu_gem_force_release(adev);
1218 amdgpu_vm_manager_fini(adev);
1219 kfree(adev->gmc.vm_fault_info);
1220 amdgpu_gart_table_vram_free(adev);
1221 amdgpu_bo_fini(adev);
1222 amdgpu_gart_fini(adev);
1223 release_firmware(adev->gmc.fw);
1224 adev->gmc.fw = NULL;
1225
1226 return 0;
1227 }
1228
gmc_v8_0_hw_init(void * handle)1229 static int gmc_v8_0_hw_init(void *handle)
1230 {
1231 int r;
1232 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1233
1234 gmc_v8_0_init_golden_registers(adev);
1235
1236 gmc_v8_0_mc_program(adev);
1237
1238 if (adev->asic_type == CHIP_TONGA) {
1239 r = gmc_v8_0_tonga_mc_load_microcode(adev);
1240 if (r) {
1241 DRM_ERROR("Failed to load MC firmware!\n");
1242 return r;
1243 }
1244 } else if (adev->asic_type == CHIP_POLARIS11 ||
1245 adev->asic_type == CHIP_POLARIS10 ||
1246 adev->asic_type == CHIP_POLARIS12) {
1247 r = gmc_v8_0_polaris_mc_load_microcode(adev);
1248 if (r) {
1249 DRM_ERROR("Failed to load MC firmware!\n");
1250 return r;
1251 }
1252 }
1253
1254 r = gmc_v8_0_gart_enable(adev);
1255 if (r)
1256 return r;
1257
1258 return r;
1259 }
1260
gmc_v8_0_hw_fini(void * handle)1261 static int gmc_v8_0_hw_fini(void *handle)
1262 {
1263 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1264
1265 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1266 gmc_v8_0_gart_disable(adev);
1267
1268 return 0;
1269 }
1270
gmc_v8_0_suspend(void * handle)1271 static int gmc_v8_0_suspend(void *handle)
1272 {
1273 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1274
1275 gmc_v8_0_hw_fini(adev);
1276
1277 return 0;
1278 }
1279
gmc_v8_0_resume(void * handle)1280 static int gmc_v8_0_resume(void *handle)
1281 {
1282 int r;
1283 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1284
1285 r = gmc_v8_0_hw_init(adev);
1286 if (r)
1287 return r;
1288
1289 amdgpu_vmid_reset_all(adev);
1290
1291 return 0;
1292 }
1293
gmc_v8_0_is_idle(void * handle)1294 static bool gmc_v8_0_is_idle(void *handle)
1295 {
1296 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1297 u32 tmp = RREG32(mmSRBM_STATUS);
1298
1299 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1300 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1301 return false;
1302
1303 return true;
1304 }
1305
gmc_v8_0_wait_for_idle(void * handle)1306 static int gmc_v8_0_wait_for_idle(void *handle)
1307 {
1308 unsigned i;
1309 u32 tmp;
1310 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1311
1312 for (i = 0; i < adev->usec_timeout; i++) {
1313 /* read MC_STATUS */
1314 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1315 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1316 SRBM_STATUS__MCC_BUSY_MASK |
1317 SRBM_STATUS__MCD_BUSY_MASK |
1318 SRBM_STATUS__VMC_BUSY_MASK |
1319 SRBM_STATUS__VMC1_BUSY_MASK);
1320 if (!tmp)
1321 return 0;
1322 udelay(1);
1323 }
1324 return -ETIMEDOUT;
1325
1326 }
1327
gmc_v8_0_check_soft_reset(void * handle)1328 static bool gmc_v8_0_check_soft_reset(void *handle)
1329 {
1330 u32 srbm_soft_reset = 0;
1331 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1332 u32 tmp = RREG32(mmSRBM_STATUS);
1333
1334 if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1335 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1336 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1337
1338 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1339 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1340 if (!(adev->flags & AMD_IS_APU))
1341 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1342 SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1343 }
1344 if (srbm_soft_reset) {
1345 adev->gmc.srbm_soft_reset = srbm_soft_reset;
1346 return true;
1347 } else {
1348 adev->gmc.srbm_soft_reset = 0;
1349 return false;
1350 }
1351 }
1352
gmc_v8_0_pre_soft_reset(void * handle)1353 static int gmc_v8_0_pre_soft_reset(void *handle)
1354 {
1355 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1356
1357 if (!adev->gmc.srbm_soft_reset)
1358 return 0;
1359
1360 gmc_v8_0_mc_stop(adev);
1361 if (gmc_v8_0_wait_for_idle(adev)) {
1362 dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1363 }
1364
1365 return 0;
1366 }
1367
gmc_v8_0_soft_reset(void * handle)1368 static int gmc_v8_0_soft_reset(void *handle)
1369 {
1370 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1371 u32 srbm_soft_reset;
1372
1373 if (!adev->gmc.srbm_soft_reset)
1374 return 0;
1375 srbm_soft_reset = adev->gmc.srbm_soft_reset;
1376
1377 if (srbm_soft_reset) {
1378 u32 tmp;
1379
1380 tmp = RREG32(mmSRBM_SOFT_RESET);
1381 tmp |= srbm_soft_reset;
1382 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1383 WREG32(mmSRBM_SOFT_RESET, tmp);
1384 tmp = RREG32(mmSRBM_SOFT_RESET);
1385
1386 udelay(50);
1387
1388 tmp &= ~srbm_soft_reset;
1389 WREG32(mmSRBM_SOFT_RESET, tmp);
1390 tmp = RREG32(mmSRBM_SOFT_RESET);
1391
1392 /* Wait a little for things to settle down */
1393 udelay(50);
1394 }
1395
1396 return 0;
1397 }
1398
gmc_v8_0_post_soft_reset(void * handle)1399 static int gmc_v8_0_post_soft_reset(void *handle)
1400 {
1401 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1402
1403 if (!adev->gmc.srbm_soft_reset)
1404 return 0;
1405
1406 gmc_v8_0_mc_resume(adev);
1407 return 0;
1408 }
1409
gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device * adev,struct amdgpu_irq_src * src,unsigned type,enum amdgpu_interrupt_state state)1410 static int gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1411 struct amdgpu_irq_src *src,
1412 unsigned type,
1413 enum amdgpu_interrupt_state state)
1414 {
1415 u32 tmp;
1416 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1417 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1418 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1419 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1420 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1421 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1422 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1423
1424 switch (state) {
1425 case AMDGPU_IRQ_STATE_DISABLE:
1426 /* system context */
1427 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1428 tmp &= ~bits;
1429 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1430 /* VMs */
1431 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1432 tmp &= ~bits;
1433 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1434 break;
1435 case AMDGPU_IRQ_STATE_ENABLE:
1436 /* system context */
1437 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1438 tmp |= bits;
1439 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1440 /* VMs */
1441 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1442 tmp |= bits;
1443 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1444 break;
1445 default:
1446 break;
1447 }
1448
1449 return 0;
1450 }
1451
gmc_v8_0_process_interrupt(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)1452 static int gmc_v8_0_process_interrupt(struct amdgpu_device *adev,
1453 struct amdgpu_irq_src *source,
1454 struct amdgpu_iv_entry *entry)
1455 {
1456 u32 addr, status, mc_client, vmid;
1457
1458 if (amdgpu_sriov_vf(adev)) {
1459 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1460 entry->src_id, entry->src_data[0]);
1461 dev_err(adev->dev, " Can't decode VM fault info here on SRIOV VF\n");
1462 return 0;
1463 }
1464
1465 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1466 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1467 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1468 /* reset addr and status */
1469 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1470
1471 if (!addr && !status)
1472 return 0;
1473
1474 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1475 gmc_v8_0_set_fault_enable_default(adev, false);
1476
1477 if (printk_ratelimit()) {
1478 struct amdgpu_task_info task_info = { 0 };
1479
1480 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
1481
1482 dev_err(adev->dev, "GPU fault detected: %d 0x%08x for process %s pid %d thread %s pid %d\n",
1483 entry->src_id, entry->src_data[0], task_info.process_name,
1484 task_info.tgid, task_info.task_name, task_info.pid);
1485 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
1486 addr);
1487 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1488 status);
1489 gmc_v8_0_vm_decode_fault(adev, status, addr, mc_client,
1490 entry->pasid);
1491 }
1492
1493 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1494 VMID);
1495 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid)
1496 && !atomic_read(&adev->gmc.vm_fault_info_updated)) {
1497 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info;
1498 u32 protections = REG_GET_FIELD(status,
1499 VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1500 PROTECTIONS);
1501
1502 info->vmid = vmid;
1503 info->mc_id = REG_GET_FIELD(status,
1504 VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1505 MEMORY_CLIENT_ID);
1506 info->status = status;
1507 info->page_addr = addr;
1508 info->prot_valid = protections & 0x7 ? true : false;
1509 info->prot_read = protections & 0x8 ? true : false;
1510 info->prot_write = protections & 0x10 ? true : false;
1511 info->prot_exec = protections & 0x20 ? true : false;
1512 mb();
1513 atomic_set(&adev->gmc.vm_fault_info_updated, 1);
1514 }
1515
1516 return 0;
1517 }
1518
fiji_update_mc_medium_grain_clock_gating(struct amdgpu_device * adev,bool enable)1519 static void fiji_update_mc_medium_grain_clock_gating(struct amdgpu_device *adev,
1520 bool enable)
1521 {
1522 uint32_t data;
1523
1524 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) {
1525 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1526 data |= MC_HUB_MISC_HUB_CG__ENABLE_MASK;
1527 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1528
1529 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1530 data |= MC_HUB_MISC_SIP_CG__ENABLE_MASK;
1531 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1532
1533 data = RREG32(mmMC_HUB_MISC_VM_CG);
1534 data |= MC_HUB_MISC_VM_CG__ENABLE_MASK;
1535 WREG32(mmMC_HUB_MISC_VM_CG, data);
1536
1537 data = RREG32(mmMC_XPB_CLK_GAT);
1538 data |= MC_XPB_CLK_GAT__ENABLE_MASK;
1539 WREG32(mmMC_XPB_CLK_GAT, data);
1540
1541 data = RREG32(mmATC_MISC_CG);
1542 data |= ATC_MISC_CG__ENABLE_MASK;
1543 WREG32(mmATC_MISC_CG, data);
1544
1545 data = RREG32(mmMC_CITF_MISC_WR_CG);
1546 data |= MC_CITF_MISC_WR_CG__ENABLE_MASK;
1547 WREG32(mmMC_CITF_MISC_WR_CG, data);
1548
1549 data = RREG32(mmMC_CITF_MISC_RD_CG);
1550 data |= MC_CITF_MISC_RD_CG__ENABLE_MASK;
1551 WREG32(mmMC_CITF_MISC_RD_CG, data);
1552
1553 data = RREG32(mmMC_CITF_MISC_VM_CG);
1554 data |= MC_CITF_MISC_VM_CG__ENABLE_MASK;
1555 WREG32(mmMC_CITF_MISC_VM_CG, data);
1556
1557 data = RREG32(mmVM_L2_CG);
1558 data |= VM_L2_CG__ENABLE_MASK;
1559 WREG32(mmVM_L2_CG, data);
1560 } else {
1561 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1562 data &= ~MC_HUB_MISC_HUB_CG__ENABLE_MASK;
1563 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1564
1565 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1566 data &= ~MC_HUB_MISC_SIP_CG__ENABLE_MASK;
1567 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1568
1569 data = RREG32(mmMC_HUB_MISC_VM_CG);
1570 data &= ~MC_HUB_MISC_VM_CG__ENABLE_MASK;
1571 WREG32(mmMC_HUB_MISC_VM_CG, data);
1572
1573 data = RREG32(mmMC_XPB_CLK_GAT);
1574 data &= ~MC_XPB_CLK_GAT__ENABLE_MASK;
1575 WREG32(mmMC_XPB_CLK_GAT, data);
1576
1577 data = RREG32(mmATC_MISC_CG);
1578 data &= ~ATC_MISC_CG__ENABLE_MASK;
1579 WREG32(mmATC_MISC_CG, data);
1580
1581 data = RREG32(mmMC_CITF_MISC_WR_CG);
1582 data &= ~MC_CITF_MISC_WR_CG__ENABLE_MASK;
1583 WREG32(mmMC_CITF_MISC_WR_CG, data);
1584
1585 data = RREG32(mmMC_CITF_MISC_RD_CG);
1586 data &= ~MC_CITF_MISC_RD_CG__ENABLE_MASK;
1587 WREG32(mmMC_CITF_MISC_RD_CG, data);
1588
1589 data = RREG32(mmMC_CITF_MISC_VM_CG);
1590 data &= ~MC_CITF_MISC_VM_CG__ENABLE_MASK;
1591 WREG32(mmMC_CITF_MISC_VM_CG, data);
1592
1593 data = RREG32(mmVM_L2_CG);
1594 data &= ~VM_L2_CG__ENABLE_MASK;
1595 WREG32(mmVM_L2_CG, data);
1596 }
1597 }
1598
fiji_update_mc_light_sleep(struct amdgpu_device * adev,bool enable)1599 static void fiji_update_mc_light_sleep(struct amdgpu_device *adev,
1600 bool enable)
1601 {
1602 uint32_t data;
1603
1604 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) {
1605 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1606 data |= MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK;
1607 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1608
1609 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1610 data |= MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK;
1611 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1612
1613 data = RREG32(mmMC_HUB_MISC_VM_CG);
1614 data |= MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1615 WREG32(mmMC_HUB_MISC_VM_CG, data);
1616
1617 data = RREG32(mmMC_XPB_CLK_GAT);
1618 data |= MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK;
1619 WREG32(mmMC_XPB_CLK_GAT, data);
1620
1621 data = RREG32(mmATC_MISC_CG);
1622 data |= ATC_MISC_CG__MEM_LS_ENABLE_MASK;
1623 WREG32(mmATC_MISC_CG, data);
1624
1625 data = RREG32(mmMC_CITF_MISC_WR_CG);
1626 data |= MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK;
1627 WREG32(mmMC_CITF_MISC_WR_CG, data);
1628
1629 data = RREG32(mmMC_CITF_MISC_RD_CG);
1630 data |= MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK;
1631 WREG32(mmMC_CITF_MISC_RD_CG, data);
1632
1633 data = RREG32(mmMC_CITF_MISC_VM_CG);
1634 data |= MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1635 WREG32(mmMC_CITF_MISC_VM_CG, data);
1636
1637 data = RREG32(mmVM_L2_CG);
1638 data |= VM_L2_CG__MEM_LS_ENABLE_MASK;
1639 WREG32(mmVM_L2_CG, data);
1640 } else {
1641 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1642 data &= ~MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK;
1643 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1644
1645 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1646 data &= ~MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK;
1647 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1648
1649 data = RREG32(mmMC_HUB_MISC_VM_CG);
1650 data &= ~MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1651 WREG32(mmMC_HUB_MISC_VM_CG, data);
1652
1653 data = RREG32(mmMC_XPB_CLK_GAT);
1654 data &= ~MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK;
1655 WREG32(mmMC_XPB_CLK_GAT, data);
1656
1657 data = RREG32(mmATC_MISC_CG);
1658 data &= ~ATC_MISC_CG__MEM_LS_ENABLE_MASK;
1659 WREG32(mmATC_MISC_CG, data);
1660
1661 data = RREG32(mmMC_CITF_MISC_WR_CG);
1662 data &= ~MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK;
1663 WREG32(mmMC_CITF_MISC_WR_CG, data);
1664
1665 data = RREG32(mmMC_CITF_MISC_RD_CG);
1666 data &= ~MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK;
1667 WREG32(mmMC_CITF_MISC_RD_CG, data);
1668
1669 data = RREG32(mmMC_CITF_MISC_VM_CG);
1670 data &= ~MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1671 WREG32(mmMC_CITF_MISC_VM_CG, data);
1672
1673 data = RREG32(mmVM_L2_CG);
1674 data &= ~VM_L2_CG__MEM_LS_ENABLE_MASK;
1675 WREG32(mmVM_L2_CG, data);
1676 }
1677 }
1678
gmc_v8_0_set_clockgating_state(void * handle,enum amd_clockgating_state state)1679 static int gmc_v8_0_set_clockgating_state(void *handle,
1680 enum amd_clockgating_state state)
1681 {
1682 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1683
1684 if (amdgpu_sriov_vf(adev))
1685 return 0;
1686
1687 switch (adev->asic_type) {
1688 case CHIP_FIJI:
1689 fiji_update_mc_medium_grain_clock_gating(adev,
1690 state == AMD_CG_STATE_GATE);
1691 fiji_update_mc_light_sleep(adev,
1692 state == AMD_CG_STATE_GATE);
1693 break;
1694 default:
1695 break;
1696 }
1697 return 0;
1698 }
1699
gmc_v8_0_set_powergating_state(void * handle,enum amd_powergating_state state)1700 static int gmc_v8_0_set_powergating_state(void *handle,
1701 enum amd_powergating_state state)
1702 {
1703 return 0;
1704 }
1705
gmc_v8_0_get_clockgating_state(void * handle,u32 * flags)1706 static void gmc_v8_0_get_clockgating_state(void *handle, u32 *flags)
1707 {
1708 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1709 int data;
1710
1711 if (amdgpu_sriov_vf(adev))
1712 *flags = 0;
1713
1714 /* AMD_CG_SUPPORT_MC_MGCG */
1715 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1716 if (data & MC_HUB_MISC_HUB_CG__ENABLE_MASK)
1717 *flags |= AMD_CG_SUPPORT_MC_MGCG;
1718
1719 /* AMD_CG_SUPPORT_MC_LS */
1720 if (data & MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK)
1721 *flags |= AMD_CG_SUPPORT_MC_LS;
1722 }
1723
1724 static const struct amd_ip_funcs gmc_v8_0_ip_funcs = {
1725 .name = "gmc_v8_0",
1726 .early_init = gmc_v8_0_early_init,
1727 .late_init = gmc_v8_0_late_init,
1728 .sw_init = gmc_v8_0_sw_init,
1729 .sw_fini = gmc_v8_0_sw_fini,
1730 .hw_init = gmc_v8_0_hw_init,
1731 .hw_fini = gmc_v8_0_hw_fini,
1732 .suspend = gmc_v8_0_suspend,
1733 .resume = gmc_v8_0_resume,
1734 .is_idle = gmc_v8_0_is_idle,
1735 .wait_for_idle = gmc_v8_0_wait_for_idle,
1736 .check_soft_reset = gmc_v8_0_check_soft_reset,
1737 .pre_soft_reset = gmc_v8_0_pre_soft_reset,
1738 .soft_reset = gmc_v8_0_soft_reset,
1739 .post_soft_reset = gmc_v8_0_post_soft_reset,
1740 .set_clockgating_state = gmc_v8_0_set_clockgating_state,
1741 .set_powergating_state = gmc_v8_0_set_powergating_state,
1742 .get_clockgating_state = gmc_v8_0_get_clockgating_state,
1743 };
1744
1745 static const struct amdgpu_gmc_funcs gmc_v8_0_gmc_funcs = {
1746 .flush_gpu_tlb = gmc_v8_0_flush_gpu_tlb,
1747 .emit_flush_gpu_tlb = gmc_v8_0_emit_flush_gpu_tlb,
1748 .emit_pasid_mapping = gmc_v8_0_emit_pasid_mapping,
1749 .set_pte_pde = gmc_v8_0_set_pte_pde,
1750 .set_prt = gmc_v8_0_set_prt,
1751 .get_vm_pte_flags = gmc_v8_0_get_vm_pte_flags,
1752 .get_vm_pde = gmc_v8_0_get_vm_pde
1753 };
1754
1755 static const struct amdgpu_irq_src_funcs gmc_v8_0_irq_funcs = {
1756 .set = gmc_v8_0_vm_fault_interrupt_state,
1757 .process = gmc_v8_0_process_interrupt,
1758 };
1759
gmc_v8_0_set_gmc_funcs(struct amdgpu_device * adev)1760 static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev)
1761 {
1762 if (adev->gmc.gmc_funcs == NULL)
1763 adev->gmc.gmc_funcs = &gmc_v8_0_gmc_funcs;
1764 }
1765
gmc_v8_0_set_irq_funcs(struct amdgpu_device * adev)1766 static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev)
1767 {
1768 adev->gmc.vm_fault.num_types = 1;
1769 adev->gmc.vm_fault.funcs = &gmc_v8_0_irq_funcs;
1770 }
1771
1772 const struct amdgpu_ip_block_version gmc_v8_0_ip_block =
1773 {
1774 .type = AMD_IP_BLOCK_TYPE_GMC,
1775 .major = 8,
1776 .minor = 0,
1777 .rev = 0,
1778 .funcs = &gmc_v8_0_ip_funcs,
1779 };
1780
1781 const struct amdgpu_ip_block_version gmc_v8_1_ip_block =
1782 {
1783 .type = AMD_IP_BLOCK_TYPE_GMC,
1784 .major = 8,
1785 .minor = 1,
1786 .rev = 0,
1787 .funcs = &gmc_v8_0_ip_funcs,
1788 };
1789
1790 const struct amdgpu_ip_block_version gmc_v8_5_ip_block =
1791 {
1792 .type = AMD_IP_BLOCK_TYPE_GMC,
1793 .major = 8,
1794 .minor = 5,
1795 .rev = 0,
1796 .funcs = &gmc_v8_0_ip_funcs,
1797 };
1798