Loading...
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
30#include "gmc/gmc_8_1_d.h"
31#include "gmc/gmc_8_1_sh_mask.h"
32
33#include "bif/bif_5_0_d.h"
34#include "bif/bif_5_0_sh_mask.h"
35
36#include "oss/oss_3_0_d.h"
37#include "oss/oss_3_0_sh_mask.h"
38
39#include "dce/dce_10_0_d.h"
40#include "dce/dce_10_0_sh_mask.h"
41
42#include "vid.h"
43#include "vi.h"
44
45#include "amdgpu_atombios.h"
46
47
48static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev);
49static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev);
50static int gmc_v8_0_wait_for_idle(void *handle);
51
52MODULE_FIRMWARE("amdgpu/tonga_mc.bin");
53MODULE_FIRMWARE("amdgpu/polaris11_mc.bin");
54MODULE_FIRMWARE("amdgpu/polaris10_mc.bin");
55MODULE_FIRMWARE("amdgpu/polaris12_mc.bin");
56
57static const u32 golden_settings_tonga_a11[] =
58{
59 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
60 mmMC_HUB_RDREQ_DMIF_LIMIT, 0x0000007f, 0x00000028,
61 mmMC_HUB_WDP_UMC, 0x00007fb6, 0x00000991,
62 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
63 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
64 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
65 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
66};
67
68static const u32 tonga_mgcg_cgcg_init[] =
69{
70 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
71};
72
73static const u32 golden_settings_fiji_a10[] =
74{
75 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
76 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
77 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
78 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
79};
80
81static const u32 fiji_mgcg_cgcg_init[] =
82{
83 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
84};
85
86static const u32 golden_settings_polaris11_a11[] =
87{
88 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
89 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
90 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
91 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
92};
93
94static const u32 golden_settings_polaris10_a11[] =
95{
96 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
97 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
98 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
99 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
100 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
101};
102
103static const u32 cz_mgcg_cgcg_init[] =
104{
105 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
106};
107
108static const u32 stoney_mgcg_cgcg_init[] =
109{
110 mmATC_MISC_CG, 0xffffffff, 0x000c0200,
111 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
112};
113
114static const u32 golden_settings_stoney_common[] =
115{
116 mmMC_HUB_RDREQ_UVD, MC_HUB_RDREQ_UVD__PRESCALE_MASK, 0x00000004,
117 mmMC_RD_GRP_OTH, MC_RD_GRP_OTH__UVD_MASK, 0x00600000
118};
119
120static void gmc_v8_0_init_golden_registers(struct amdgpu_device *adev)
121{
122 switch (adev->asic_type) {
123 case CHIP_FIJI:
124 amdgpu_device_program_register_sequence(adev,
125 fiji_mgcg_cgcg_init,
126 ARRAY_SIZE(fiji_mgcg_cgcg_init));
127 amdgpu_device_program_register_sequence(adev,
128 golden_settings_fiji_a10,
129 ARRAY_SIZE(golden_settings_fiji_a10));
130 break;
131 case CHIP_TONGA:
132 amdgpu_device_program_register_sequence(adev,
133 tonga_mgcg_cgcg_init,
134 ARRAY_SIZE(tonga_mgcg_cgcg_init));
135 amdgpu_device_program_register_sequence(adev,
136 golden_settings_tonga_a11,
137 ARRAY_SIZE(golden_settings_tonga_a11));
138 break;
139 case CHIP_POLARIS11:
140 case CHIP_POLARIS12:
141 amdgpu_device_program_register_sequence(adev,
142 golden_settings_polaris11_a11,
143 ARRAY_SIZE(golden_settings_polaris11_a11));
144 break;
145 case CHIP_POLARIS10:
146 amdgpu_device_program_register_sequence(adev,
147 golden_settings_polaris10_a11,
148 ARRAY_SIZE(golden_settings_polaris10_a11));
149 break;
150 case CHIP_CARRIZO:
151 amdgpu_device_program_register_sequence(adev,
152 cz_mgcg_cgcg_init,
153 ARRAY_SIZE(cz_mgcg_cgcg_init));
154 break;
155 case CHIP_STONEY:
156 amdgpu_device_program_register_sequence(adev,
157 stoney_mgcg_cgcg_init,
158 ARRAY_SIZE(stoney_mgcg_cgcg_init));
159 amdgpu_device_program_register_sequence(adev,
160 golden_settings_stoney_common,
161 ARRAY_SIZE(golden_settings_stoney_common));
162 break;
163 default:
164 break;
165 }
166}
167
168static void gmc_v8_0_mc_stop(struct amdgpu_device *adev)
169{
170 u32 blackout;
171
172 gmc_v8_0_wait_for_idle(adev);
173
174 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
175 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) {
176 /* Block CPU access */
177 WREG32(mmBIF_FB_EN, 0);
178 /* blackout the MC */
179 blackout = REG_SET_FIELD(blackout,
180 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 1);
181 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout);
182 }
183 /* wait for the MC to settle */
184 udelay(100);
185}
186
187static void gmc_v8_0_mc_resume(struct amdgpu_device *adev)
188{
189 u32 tmp;
190
191 /* unblackout the MC */
192 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
193 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
194 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp);
195 /* allow CPU access */
196 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1);
197 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1);
198 WREG32(mmBIF_FB_EN, tmp);
199}
200
201/**
202 * gmc_v8_0_init_microcode - load ucode images from disk
203 *
204 * @adev: amdgpu_device pointer
205 *
206 * Use the firmware interface to load the ucode images into
207 * the driver (not loaded into hw).
208 * Returns 0 on success, error on failure.
209 */
210static int gmc_v8_0_init_microcode(struct amdgpu_device *adev)
211{
212 const char *chip_name;
213 char fw_name[30];
214 int err;
215
216 DRM_DEBUG("\n");
217
218 switch (adev->asic_type) {
219 case CHIP_TONGA:
220 chip_name = "tonga";
221 break;
222 case CHIP_POLARIS11:
223 chip_name = "polaris11";
224 break;
225 case CHIP_POLARIS10:
226 chip_name = "polaris10";
227 break;
228 case CHIP_POLARIS12:
229 chip_name = "polaris12";
230 break;
231 case CHIP_FIJI:
232 case CHIP_CARRIZO:
233 case CHIP_STONEY:
234 return 0;
235 default: BUG();
236 }
237
238 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name);
239 err = request_firmware(&adev->gmc.fw, fw_name, adev->dev);
240 if (err)
241 goto out;
242 err = amdgpu_ucode_validate(adev->gmc.fw);
243
244out:
245 if (err) {
246 pr_err("mc: Failed to load firmware \"%s\"\n", fw_name);
247 release_firmware(adev->gmc.fw);
248 adev->gmc.fw = NULL;
249 }
250 return err;
251}
252
253/**
254 * gmc_v8_0_tonga_mc_load_microcode - load tonga MC ucode into the hw
255 *
256 * @adev: amdgpu_device pointer
257 *
258 * Load the GDDR MC ucode into the hw (CIK).
259 * Returns 0 on success, error on failure.
260 */
261static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
262{
263 const struct mc_firmware_header_v1_0 *hdr;
264 const __le32 *fw_data = NULL;
265 const __le32 *io_mc_regs = NULL;
266 u32 running;
267 int i, ucode_size, regs_size;
268
269 /* Skip MC ucode loading on SR-IOV capable boards.
270 * vbios does this for us in asic_init in that case.
271 * Skip MC ucode loading on VF, because hypervisor will do that
272 * for this adaptor.
273 */
274 if (amdgpu_sriov_bios(adev))
275 return 0;
276
277 if (!adev->gmc.fw)
278 return -EINVAL;
279
280 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
281 amdgpu_ucode_print_mc_hdr(&hdr->header);
282
283 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
284 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
285 io_mc_regs = (const __le32 *)
286 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
287 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
288 fw_data = (const __le32 *)
289 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
290
291 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN);
292
293 if (running == 0) {
294 /* reset the engine and set to writable */
295 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
296 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
297
298 /* load mc io regs */
299 for (i = 0; i < regs_size; i++) {
300 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
301 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
302 }
303 /* load the MC ucode */
304 for (i = 0; i < ucode_size; i++)
305 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
306
307 /* put the engine back into the active state */
308 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
309 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
310 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
311
312 /* wait for training to complete */
313 for (i = 0; i < adev->usec_timeout; i++) {
314 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
315 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0))
316 break;
317 udelay(1);
318 }
319 for (i = 0; i < adev->usec_timeout; i++) {
320 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
321 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1))
322 break;
323 udelay(1);
324 }
325 }
326
327 return 0;
328}
329
330static int gmc_v8_0_polaris_mc_load_microcode(struct amdgpu_device *adev)
331{
332 const struct mc_firmware_header_v1_0 *hdr;
333 const __le32 *fw_data = NULL;
334 const __le32 *io_mc_regs = NULL;
335 u32 data, vbios_version;
336 int i, ucode_size, regs_size;
337
338 /* Skip MC ucode loading on SR-IOV capable boards.
339 * vbios does this for us in asic_init in that case.
340 * Skip MC ucode loading on VF, because hypervisor will do that
341 * for this adaptor.
342 */
343 if (amdgpu_sriov_bios(adev))
344 return 0;
345
346 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, 0x9F);
347 data = RREG32(mmMC_SEQ_IO_DEBUG_DATA);
348 vbios_version = data & 0xf;
349
350 if (vbios_version == 0)
351 return 0;
352
353 if (!adev->gmc.fw)
354 return -EINVAL;
355
356 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
357 amdgpu_ucode_print_mc_hdr(&hdr->header);
358
359 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
360 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
361 io_mc_regs = (const __le32 *)
362 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
363 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
364 fw_data = (const __le32 *)
365 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
366
367 data = RREG32(mmMC_SEQ_MISC0);
368 data &= ~(0x40);
369 WREG32(mmMC_SEQ_MISC0, data);
370
371 /* load mc io regs */
372 for (i = 0; i < regs_size; i++) {
373 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
374 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
375 }
376
377 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
378 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
379
380 /* load the MC ucode */
381 for (i = 0; i < ucode_size; i++)
382 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
383
384 /* put the engine back into the active state */
385 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
386 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
387 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
388
389 /* wait for training to complete */
390 for (i = 0; i < adev->usec_timeout; i++) {
391 data = RREG32(mmMC_SEQ_MISC0);
392 if (data & 0x80)
393 break;
394 udelay(1);
395 }
396
397 return 0;
398}
399
400static void gmc_v8_0_vram_gtt_location(struct amdgpu_device *adev,
401 struct amdgpu_gmc *mc)
402{
403 u64 base = 0;
404
405 if (!amdgpu_sriov_vf(adev))
406 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF;
407 base <<= 24;
408
409 amdgpu_device_vram_location(adev, &adev->gmc, base);
410 amdgpu_device_gart_location(adev, mc);
411}
412
413/**
414 * gmc_v8_0_mc_program - program the GPU memory controller
415 *
416 * @adev: amdgpu_device pointer
417 *
418 * Set the location of vram, gart, and AGP in the GPU's
419 * physical address space (CIK).
420 */
421static void gmc_v8_0_mc_program(struct amdgpu_device *adev)
422{
423 u32 tmp;
424 int i, j;
425
426 /* Initialize HDP */
427 for (i = 0, j = 0; i < 32; i++, j += 0x6) {
428 WREG32((0xb05 + j), 0x00000000);
429 WREG32((0xb06 + j), 0x00000000);
430 WREG32((0xb07 + j), 0x00000000);
431 WREG32((0xb08 + j), 0x00000000);
432 WREG32((0xb09 + j), 0x00000000);
433 }
434 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0);
435
436 if (gmc_v8_0_wait_for_idle((void *)adev)) {
437 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
438 }
439 if (adev->mode_info.num_crtc) {
440 /* Lockout access through VGA aperture*/
441 tmp = RREG32(mmVGA_HDP_CONTROL);
442 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
443 WREG32(mmVGA_HDP_CONTROL, tmp);
444
445 /* disable VGA render */
446 tmp = RREG32(mmVGA_RENDER_CONTROL);
447 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
448 WREG32(mmVGA_RENDER_CONTROL, tmp);
449 }
450 /* Update configuration */
451 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR,
452 adev->gmc.vram_start >> 12);
453 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR,
454 adev->gmc.vram_end >> 12);
455 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
456 adev->vram_scratch.gpu_addr >> 12);
457
458 if (amdgpu_sriov_vf(adev)) {
459 tmp = ((adev->gmc.vram_end >> 24) & 0xFFFF) << 16;
460 tmp |= ((adev->gmc.vram_start >> 24) & 0xFFFF);
461 WREG32(mmMC_VM_FB_LOCATION, tmp);
462 /* XXX double check these! */
463 WREG32(mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
464 WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
465 WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF);
466 }
467
468 WREG32(mmMC_VM_AGP_BASE, 0);
469 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF);
470 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF);
471 if (gmc_v8_0_wait_for_idle((void *)adev)) {
472 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
473 }
474
475 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
476
477 tmp = RREG32(mmHDP_MISC_CNTL);
478 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0);
479 WREG32(mmHDP_MISC_CNTL, tmp);
480
481 tmp = RREG32(mmHDP_HOST_PATH_CNTL);
482 WREG32(mmHDP_HOST_PATH_CNTL, tmp);
483}
484
485/**
486 * gmc_v8_0_mc_init - initialize the memory controller driver params
487 *
488 * @adev: amdgpu_device pointer
489 *
490 * Look up the amount of vram, vram width, and decide how to place
491 * vram and gart within the GPU's physical address space (CIK).
492 * Returns 0 for success.
493 */
494static int gmc_v8_0_mc_init(struct amdgpu_device *adev)
495{
496 int r;
497
498 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev);
499 if (!adev->gmc.vram_width) {
500 u32 tmp;
501 int chansize, numchan;
502
503 /* Get VRAM informations */
504 tmp = RREG32(mmMC_ARB_RAMCFG);
505 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) {
506 chansize = 64;
507 } else {
508 chansize = 32;
509 }
510 tmp = RREG32(mmMC_SHARED_CHMAP);
511 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) {
512 case 0:
513 default:
514 numchan = 1;
515 break;
516 case 1:
517 numchan = 2;
518 break;
519 case 2:
520 numchan = 4;
521 break;
522 case 3:
523 numchan = 8;
524 break;
525 case 4:
526 numchan = 3;
527 break;
528 case 5:
529 numchan = 6;
530 break;
531 case 6:
532 numchan = 10;
533 break;
534 case 7:
535 numchan = 12;
536 break;
537 case 8:
538 numchan = 16;
539 break;
540 }
541 adev->gmc.vram_width = numchan * chansize;
542 }
543 /* size in MB on si */
544 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
545 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
546
547 if (!(adev->flags & AMD_IS_APU)) {
548 r = amdgpu_device_resize_fb_bar(adev);
549 if (r)
550 return r;
551 }
552 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
553 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
554
555#ifdef CONFIG_X86_64
556 if (adev->flags & AMD_IS_APU) {
557 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22;
558 adev->gmc.aper_size = adev->gmc.real_vram_size;
559 }
560#endif
561
562 /* In case the PCI BAR is larger than the actual amount of vram */
563 adev->gmc.visible_vram_size = adev->gmc.aper_size;
564 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
565 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
566
567 /* set the gart size */
568 if (amdgpu_gart_size == -1) {
569 switch (adev->asic_type) {
570 case CHIP_POLARIS11: /* all engines support GPUVM */
571 case CHIP_POLARIS10: /* all engines support GPUVM */
572 case CHIP_POLARIS12: /* all engines support GPUVM */
573 default:
574 adev->gmc.gart_size = 256ULL << 20;
575 break;
576 case CHIP_TONGA: /* UVD, VCE do not support GPUVM */
577 case CHIP_FIJI: /* UVD, VCE do not support GPUVM */
578 case CHIP_CARRIZO: /* UVD, VCE do not support GPUVM, DCE SG support */
579 case CHIP_STONEY: /* UVD does not support GPUVM, DCE SG support */
580 adev->gmc.gart_size = 1024ULL << 20;
581 break;
582 }
583 } else {
584 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
585 }
586
587 gmc_v8_0_vram_gtt_location(adev, &adev->gmc);
588
589 return 0;
590}
591
592/*
593 * GART
594 * VMID 0 is the physical GPU addresses as used by the kernel.
595 * VMIDs 1-15 are used for userspace clients and are handled
596 * by the amdgpu vm/hsa code.
597 */
598
599/**
600 * gmc_v8_0_flush_gpu_tlb - gart tlb flush callback
601 *
602 * @adev: amdgpu_device pointer
603 * @vmid: vm instance to flush
604 *
605 * Flush the TLB for the requested page table (CIK).
606 */
607static void gmc_v8_0_flush_gpu_tlb(struct amdgpu_device *adev,
608 uint32_t vmid)
609{
610 /* bits 0-15 are the VM contexts0-15 */
611 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
612}
613
614static uint64_t gmc_v8_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
615 unsigned vmid, uint64_t pd_addr)
616{
617 uint32_t reg;
618
619 if (vmid < 8)
620 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid;
621 else
622 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8;
623 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12);
624
625 /* bits 0-15 are the VM contexts0-15 */
626 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid);
627
628 return pd_addr;
629}
630
631static void gmc_v8_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
632 unsigned pasid)
633{
634 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid);
635}
636
637/**
638 * gmc_v8_0_set_pte_pde - update the page tables using MMIO
639 *
640 * @adev: amdgpu_device pointer
641 * @cpu_pt_addr: cpu address of the page table
642 * @gpu_page_idx: entry in the page table to update
643 * @addr: dst addr to write into pte/pde
644 * @flags: access flags
645 *
646 * Update the page tables using the CPU.
647 */
648static int gmc_v8_0_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
649 uint32_t gpu_page_idx, uint64_t addr,
650 uint64_t flags)
651{
652 void __iomem *ptr = (void *)cpu_pt_addr;
653 uint64_t value;
654
655 /*
656 * PTE format on VI:
657 * 63:40 reserved
658 * 39:12 4k physical page base address
659 * 11:7 fragment
660 * 6 write
661 * 5 read
662 * 4 exe
663 * 3 reserved
664 * 2 snooped
665 * 1 system
666 * 0 valid
667 *
668 * PDE format on VI:
669 * 63:59 block fragment size
670 * 58:40 reserved
671 * 39:1 physical base address of PTE
672 * bits 5:1 must be 0.
673 * 0 valid
674 */
675 value = addr & 0x000000FFFFFFF000ULL;
676 value |= flags;
677 writeq(value, ptr + (gpu_page_idx * 8));
678
679 return 0;
680}
681
682static uint64_t gmc_v8_0_get_vm_pte_flags(struct amdgpu_device *adev,
683 uint32_t flags)
684{
685 uint64_t pte_flag = 0;
686
687 if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
688 pte_flag |= AMDGPU_PTE_EXECUTABLE;
689 if (flags & AMDGPU_VM_PAGE_READABLE)
690 pte_flag |= AMDGPU_PTE_READABLE;
691 if (flags & AMDGPU_VM_PAGE_WRITEABLE)
692 pte_flag |= AMDGPU_PTE_WRITEABLE;
693 if (flags & AMDGPU_VM_PAGE_PRT)
694 pte_flag |= AMDGPU_PTE_PRT;
695
696 return pte_flag;
697}
698
699static void gmc_v8_0_get_vm_pde(struct amdgpu_device *adev, int level,
700 uint64_t *addr, uint64_t *flags)
701{
702 BUG_ON(*addr & 0xFFFFFF0000000FFFULL);
703}
704
705/**
706 * gmc_v8_0_set_fault_enable_default - update VM fault handling
707 *
708 * @adev: amdgpu_device pointer
709 * @value: true redirects VM faults to the default page
710 */
711static void gmc_v8_0_set_fault_enable_default(struct amdgpu_device *adev,
712 bool value)
713{
714 u32 tmp;
715
716 tmp = RREG32(mmVM_CONTEXT1_CNTL);
717 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
718 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
719 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
720 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
721 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
722 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
723 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
724 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
725 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
726 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
727 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
728 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
729 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
730 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
731 WREG32(mmVM_CONTEXT1_CNTL, tmp);
732}
733
734/**
735 * gmc_v8_0_set_prt - set PRT VM fault
736 *
737 * @adev: amdgpu_device pointer
738 * @enable: enable/disable VM fault handling for PRT
739*/
740static void gmc_v8_0_set_prt(struct amdgpu_device *adev, bool enable)
741{
742 u32 tmp;
743
744 if (enable && !adev->gmc.prt_warning) {
745 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n");
746 adev->gmc.prt_warning = true;
747 }
748
749 tmp = RREG32(mmVM_PRT_CNTL);
750 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
751 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
752 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
753 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
754 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
755 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
756 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
757 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
758 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
759 L2_CACHE_STORE_INVALID_ENTRIES, enable);
760 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
761 L1_TLB_STORE_INVALID_ENTRIES, enable);
762 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
763 MASK_PDE0_FAULT, enable);
764 WREG32(mmVM_PRT_CNTL, tmp);
765
766 if (enable) {
767 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT;
768 uint32_t high = adev->vm_manager.max_pfn -
769 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT);
770
771 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low);
772 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low);
773 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low);
774 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low);
775 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high);
776 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high);
777 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high);
778 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high);
779 } else {
780 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff);
781 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff);
782 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff);
783 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff);
784 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0);
785 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0);
786 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0);
787 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0);
788 }
789}
790
791/**
792 * gmc_v8_0_gart_enable - gart enable
793 *
794 * @adev: amdgpu_device pointer
795 *
796 * This sets up the TLBs, programs the page tables for VMID0,
797 * sets up the hw for VMIDs 1-15 which are allocated on
798 * demand, and sets up the global locations for the LDS, GDS,
799 * and GPUVM for FSA64 clients (CIK).
800 * Returns 0 for success, errors for failure.
801 */
802static int gmc_v8_0_gart_enable(struct amdgpu_device *adev)
803{
804 int r, i;
805 u32 tmp, field;
806
807 if (adev->gart.robj == NULL) {
808 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
809 return -EINVAL;
810 }
811 r = amdgpu_gart_table_vram_pin(adev);
812 if (r)
813 return r;
814 /* Setup TLB control */
815 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
816 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
817 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
818 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
819 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
820 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
821 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
822 /* Setup L2 cache */
823 tmp = RREG32(mmVM_L2_CNTL);
824 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
825 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
826 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
827 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
828 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
829 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
830 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1);
831 WREG32(mmVM_L2_CNTL, tmp);
832 tmp = RREG32(mmVM_L2_CNTL2);
833 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
834 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
835 WREG32(mmVM_L2_CNTL2, tmp);
836
837 field = adev->vm_manager.fragment_size;
838 tmp = RREG32(mmVM_L2_CNTL3);
839 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
840 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field);
841 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field);
842 WREG32(mmVM_L2_CNTL3, tmp);
843 /* XXX: set to enable PTE/PDE in system memory */
844 tmp = RREG32(mmVM_L2_CNTL4);
845 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_PHYSICAL, 0);
846 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SHARED, 0);
847 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SNOOP, 0);
848 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_PHYSICAL, 0);
849 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SHARED, 0);
850 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SNOOP, 0);
851 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_PHYSICAL, 0);
852 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SHARED, 0);
853 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SNOOP, 0);
854 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_PHYSICAL, 0);
855 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SHARED, 0);
856 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SNOOP, 0);
857 WREG32(mmVM_L2_CNTL4, tmp);
858 /* setup context0 */
859 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12);
860 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12);
861 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, adev->gart.table_addr >> 12);
862 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
863 (u32)(adev->dummy_page_addr >> 12));
864 WREG32(mmVM_CONTEXT0_CNTL2, 0);
865 tmp = RREG32(mmVM_CONTEXT0_CNTL);
866 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
867 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
868 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
869 WREG32(mmVM_CONTEXT0_CNTL, tmp);
870
871 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR, 0);
872 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR, 0);
873 WREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET, 0);
874
875 /* empty context1-15 */
876 /* FIXME start with 4G, once using 2 level pt switch to full
877 * vm size space
878 */
879 /* set vm size, must be a multiple of 4 */
880 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
881 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
882 for (i = 1; i < 16; i++) {
883 if (i < 8)
884 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
885 adev->gart.table_addr >> 12);
886 else
887 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
888 adev->gart.table_addr >> 12);
889 }
890
891 /* enable context1-15 */
892 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
893 (u32)(adev->dummy_page_addr >> 12));
894 WREG32(mmVM_CONTEXT1_CNTL2, 4);
895 tmp = RREG32(mmVM_CONTEXT1_CNTL);
896 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
897 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
898 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
899 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
900 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
901 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
902 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
903 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
904 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
905 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
906 adev->vm_manager.block_size - 9);
907 WREG32(mmVM_CONTEXT1_CNTL, tmp);
908 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
909 gmc_v8_0_set_fault_enable_default(adev, false);
910 else
911 gmc_v8_0_set_fault_enable_default(adev, true);
912
913 gmc_v8_0_flush_gpu_tlb(adev, 0);
914 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
915 (unsigned)(adev->gmc.gart_size >> 20),
916 (unsigned long long)adev->gart.table_addr);
917 adev->gart.ready = true;
918 return 0;
919}
920
921static int gmc_v8_0_gart_init(struct amdgpu_device *adev)
922{
923 int r;
924
925 if (adev->gart.robj) {
926 WARN(1, "R600 PCIE GART already initialized\n");
927 return 0;
928 }
929 /* Initialize common gart structure */
930 r = amdgpu_gart_init(adev);
931 if (r)
932 return r;
933 adev->gart.table_size = adev->gart.num_gpu_pages * 8;
934 adev->gart.gart_pte_flags = AMDGPU_PTE_EXECUTABLE;
935 return amdgpu_gart_table_vram_alloc(adev);
936}
937
938/**
939 * gmc_v8_0_gart_disable - gart disable
940 *
941 * @adev: amdgpu_device pointer
942 *
943 * This disables all VM page table (CIK).
944 */
945static void gmc_v8_0_gart_disable(struct amdgpu_device *adev)
946{
947 u32 tmp;
948
949 /* Disable all tables */
950 WREG32(mmVM_CONTEXT0_CNTL, 0);
951 WREG32(mmVM_CONTEXT1_CNTL, 0);
952 /* Setup TLB control */
953 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
954 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
955 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
956 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
957 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
958 /* Setup L2 cache */
959 tmp = RREG32(mmVM_L2_CNTL);
960 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
961 WREG32(mmVM_L2_CNTL, tmp);
962 WREG32(mmVM_L2_CNTL2, 0);
963 amdgpu_gart_table_vram_unpin(adev);
964}
965
966/**
967 * gmc_v8_0_gart_fini - vm fini callback
968 *
969 * @adev: amdgpu_device pointer
970 *
971 * Tears down the driver GART/VM setup (CIK).
972 */
973static void gmc_v8_0_gart_fini(struct amdgpu_device *adev)
974{
975 amdgpu_gart_table_vram_free(adev);
976 amdgpu_gart_fini(adev);
977}
978
979/**
980 * gmc_v8_0_vm_decode_fault - print human readable fault info
981 *
982 * @adev: amdgpu_device pointer
983 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
984 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
985 *
986 * Print human readable fault information (CIK).
987 */
988static void gmc_v8_0_vm_decode_fault(struct amdgpu_device *adev, u32 status,
989 u32 addr, u32 mc_client, unsigned pasid)
990{
991 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
992 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
993 PROTECTIONS);
994 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
995 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
996 u32 mc_id;
997
998 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
999 MEMORY_CLIENT_ID);
1000
1001 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
1002 protections, vmid, pasid, addr,
1003 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1004 MEMORY_CLIENT_RW) ?
1005 "write" : "read", block, mc_client, mc_id);
1006}
1007
1008static int gmc_v8_0_convert_vram_type(int mc_seq_vram_type)
1009{
1010 switch (mc_seq_vram_type) {
1011 case MC_SEQ_MISC0__MT__GDDR1:
1012 return AMDGPU_VRAM_TYPE_GDDR1;
1013 case MC_SEQ_MISC0__MT__DDR2:
1014 return AMDGPU_VRAM_TYPE_DDR2;
1015 case MC_SEQ_MISC0__MT__GDDR3:
1016 return AMDGPU_VRAM_TYPE_GDDR3;
1017 case MC_SEQ_MISC0__MT__GDDR4:
1018 return AMDGPU_VRAM_TYPE_GDDR4;
1019 case MC_SEQ_MISC0__MT__GDDR5:
1020 return AMDGPU_VRAM_TYPE_GDDR5;
1021 case MC_SEQ_MISC0__MT__HBM:
1022 return AMDGPU_VRAM_TYPE_HBM;
1023 case MC_SEQ_MISC0__MT__DDR3:
1024 return AMDGPU_VRAM_TYPE_DDR3;
1025 default:
1026 return AMDGPU_VRAM_TYPE_UNKNOWN;
1027 }
1028}
1029
1030static int gmc_v8_0_early_init(void *handle)
1031{
1032 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1033
1034 gmc_v8_0_set_gmc_funcs(adev);
1035 gmc_v8_0_set_irq_funcs(adev);
1036
1037 adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1038 adev->gmc.shared_aperture_end =
1039 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1040 adev->gmc.private_aperture_start =
1041 adev->gmc.shared_aperture_end + 1;
1042 adev->gmc.private_aperture_end =
1043 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1044
1045 return 0;
1046}
1047
1048static int gmc_v8_0_late_init(void *handle)
1049{
1050 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1051
1052 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS)
1053 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1054 else
1055 return 0;
1056}
1057
1058#define mmMC_SEQ_MISC0_FIJI 0xA71
1059
1060static int gmc_v8_0_sw_init(void *handle)
1061{
1062 int r;
1063 int dma_bits;
1064 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1065
1066 if (adev->flags & AMD_IS_APU) {
1067 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
1068 } else {
1069 u32 tmp;
1070
1071 if (adev->asic_type == CHIP_FIJI)
1072 tmp = RREG32(mmMC_SEQ_MISC0_FIJI);
1073 else
1074 tmp = RREG32(mmMC_SEQ_MISC0);
1075 tmp &= MC_SEQ_MISC0__MT__MASK;
1076 adev->gmc.vram_type = gmc_v8_0_convert_vram_type(tmp);
1077 }
1078
1079 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 146, &adev->gmc.vm_fault);
1080 if (r)
1081 return r;
1082
1083 r = amdgpu_irq_add_id(adev, AMDGPU_IH_CLIENTID_LEGACY, 147, &adev->gmc.vm_fault);
1084 if (r)
1085 return r;
1086
1087 /* Adjust VM size here.
1088 * Currently set to 4GB ((1 << 20) 4k pages).
1089 * Max GPUVM size for cayman and SI is 40 bits.
1090 */
1091 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40);
1092
1093 /* Set the internal MC address mask
1094 * This is the max address of the GPU's
1095 * internal address space.
1096 */
1097 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
1098
1099 adev->gmc.stolen_size = 256 * 1024;
1100
1101 /* set DMA mask + need_dma32 flags.
1102 * PCIE - can handle 40-bits.
1103 * IGP - can handle 40-bits
1104 * PCI - dma32 for legacy pci gart, 40 bits on newer asics
1105 */
1106 adev->need_dma32 = false;
1107 dma_bits = adev->need_dma32 ? 32 : 40;
1108 r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
1109 if (r) {
1110 adev->need_dma32 = true;
1111 dma_bits = 32;
1112 pr_warn("amdgpu: No suitable DMA available\n");
1113 }
1114 r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
1115 if (r) {
1116 pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32));
1117 pr_warn("amdgpu: No coherent DMA available\n");
1118 }
1119 adev->need_swiotlb = drm_get_max_iomem() > ((u64)1 << dma_bits);
1120
1121 r = gmc_v8_0_init_microcode(adev);
1122 if (r) {
1123 DRM_ERROR("Failed to load mc firmware!\n");
1124 return r;
1125 }
1126
1127 r = gmc_v8_0_mc_init(adev);
1128 if (r)
1129 return r;
1130
1131 /* Memory manager */
1132 r = amdgpu_bo_init(adev);
1133 if (r)
1134 return r;
1135
1136 r = gmc_v8_0_gart_init(adev);
1137 if (r)
1138 return r;
1139
1140 /*
1141 * number of VMs
1142 * VMID 0 is reserved for System
1143 * amdgpu graphics/compute will use VMIDs 1-7
1144 * amdkfd will use VMIDs 8-15
1145 */
1146 adev->vm_manager.id_mgr[0].num_ids = AMDGPU_NUM_OF_VMIDS;
1147 amdgpu_vm_manager_init(adev);
1148
1149 /* base offset of vram pages */
1150 if (adev->flags & AMD_IS_APU) {
1151 u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
1152
1153 tmp <<= 22;
1154 adev->vm_manager.vram_base_offset = tmp;
1155 } else {
1156 adev->vm_manager.vram_base_offset = 0;
1157 }
1158
1159 return 0;
1160}
1161
1162static int gmc_v8_0_sw_fini(void *handle)
1163{
1164 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1165
1166 amdgpu_gem_force_release(adev);
1167 amdgpu_vm_manager_fini(adev);
1168 gmc_v8_0_gart_fini(adev);
1169 amdgpu_bo_fini(adev);
1170 release_firmware(adev->gmc.fw);
1171 adev->gmc.fw = NULL;
1172
1173 return 0;
1174}
1175
1176static int gmc_v8_0_hw_init(void *handle)
1177{
1178 int r;
1179 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1180
1181 gmc_v8_0_init_golden_registers(adev);
1182
1183 gmc_v8_0_mc_program(adev);
1184
1185 if (adev->asic_type == CHIP_TONGA) {
1186 r = gmc_v8_0_tonga_mc_load_microcode(adev);
1187 if (r) {
1188 DRM_ERROR("Failed to load MC firmware!\n");
1189 return r;
1190 }
1191 } else if (adev->asic_type == CHIP_POLARIS11 ||
1192 adev->asic_type == CHIP_POLARIS10 ||
1193 adev->asic_type == CHIP_POLARIS12) {
1194 r = gmc_v8_0_polaris_mc_load_microcode(adev);
1195 if (r) {
1196 DRM_ERROR("Failed to load MC firmware!\n");
1197 return r;
1198 }
1199 }
1200
1201 r = gmc_v8_0_gart_enable(adev);
1202 if (r)
1203 return r;
1204
1205 return r;
1206}
1207
1208static int gmc_v8_0_hw_fini(void *handle)
1209{
1210 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1211
1212 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1213 gmc_v8_0_gart_disable(adev);
1214
1215 return 0;
1216}
1217
1218static int gmc_v8_0_suspend(void *handle)
1219{
1220 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1221
1222 gmc_v8_0_hw_fini(adev);
1223
1224 return 0;
1225}
1226
1227static int gmc_v8_0_resume(void *handle)
1228{
1229 int r;
1230 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1231
1232 r = gmc_v8_0_hw_init(adev);
1233 if (r)
1234 return r;
1235
1236 amdgpu_vmid_reset_all(adev);
1237
1238 return 0;
1239}
1240
1241static bool gmc_v8_0_is_idle(void *handle)
1242{
1243 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1244 u32 tmp = RREG32(mmSRBM_STATUS);
1245
1246 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1247 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1248 return false;
1249
1250 return true;
1251}
1252
1253static int gmc_v8_0_wait_for_idle(void *handle)
1254{
1255 unsigned i;
1256 u32 tmp;
1257 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1258
1259 for (i = 0; i < adev->usec_timeout; i++) {
1260 /* read MC_STATUS */
1261 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1262 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1263 SRBM_STATUS__MCC_BUSY_MASK |
1264 SRBM_STATUS__MCD_BUSY_MASK |
1265 SRBM_STATUS__VMC_BUSY_MASK |
1266 SRBM_STATUS__VMC1_BUSY_MASK);
1267 if (!tmp)
1268 return 0;
1269 udelay(1);
1270 }
1271 return -ETIMEDOUT;
1272
1273}
1274
1275static bool gmc_v8_0_check_soft_reset(void *handle)
1276{
1277 u32 srbm_soft_reset = 0;
1278 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1279 u32 tmp = RREG32(mmSRBM_STATUS);
1280
1281 if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1282 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1283 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1284
1285 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1286 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1287 if (!(adev->flags & AMD_IS_APU))
1288 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1289 SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1290 }
1291 if (srbm_soft_reset) {
1292 adev->gmc.srbm_soft_reset = srbm_soft_reset;
1293 return true;
1294 } else {
1295 adev->gmc.srbm_soft_reset = 0;
1296 return false;
1297 }
1298}
1299
1300static int gmc_v8_0_pre_soft_reset(void *handle)
1301{
1302 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1303
1304 if (!adev->gmc.srbm_soft_reset)
1305 return 0;
1306
1307 gmc_v8_0_mc_stop(adev);
1308 if (gmc_v8_0_wait_for_idle(adev)) {
1309 dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1310 }
1311
1312 return 0;
1313}
1314
1315static int gmc_v8_0_soft_reset(void *handle)
1316{
1317 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1318 u32 srbm_soft_reset;
1319
1320 if (!adev->gmc.srbm_soft_reset)
1321 return 0;
1322 srbm_soft_reset = adev->gmc.srbm_soft_reset;
1323
1324 if (srbm_soft_reset) {
1325 u32 tmp;
1326
1327 tmp = RREG32(mmSRBM_SOFT_RESET);
1328 tmp |= srbm_soft_reset;
1329 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1330 WREG32(mmSRBM_SOFT_RESET, tmp);
1331 tmp = RREG32(mmSRBM_SOFT_RESET);
1332
1333 udelay(50);
1334
1335 tmp &= ~srbm_soft_reset;
1336 WREG32(mmSRBM_SOFT_RESET, tmp);
1337 tmp = RREG32(mmSRBM_SOFT_RESET);
1338
1339 /* Wait a little for things to settle down */
1340 udelay(50);
1341 }
1342
1343 return 0;
1344}
1345
1346static int gmc_v8_0_post_soft_reset(void *handle)
1347{
1348 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1349
1350 if (!adev->gmc.srbm_soft_reset)
1351 return 0;
1352
1353 gmc_v8_0_mc_resume(adev);
1354 return 0;
1355}
1356
1357static int gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1358 struct amdgpu_irq_src *src,
1359 unsigned type,
1360 enum amdgpu_interrupt_state state)
1361{
1362 u32 tmp;
1363 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1364 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1365 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1366 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1367 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1368 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1369 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1370
1371 switch (state) {
1372 case AMDGPU_IRQ_STATE_DISABLE:
1373 /* system context */
1374 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1375 tmp &= ~bits;
1376 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1377 /* VMs */
1378 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1379 tmp &= ~bits;
1380 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1381 break;
1382 case AMDGPU_IRQ_STATE_ENABLE:
1383 /* system context */
1384 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1385 tmp |= bits;
1386 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1387 /* VMs */
1388 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1389 tmp |= bits;
1390 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1391 break;
1392 default:
1393 break;
1394 }
1395
1396 return 0;
1397}
1398
1399static int gmc_v8_0_process_interrupt(struct amdgpu_device *adev,
1400 struct amdgpu_irq_src *source,
1401 struct amdgpu_iv_entry *entry)
1402{
1403 u32 addr, status, mc_client;
1404
1405 if (amdgpu_sriov_vf(adev)) {
1406 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1407 entry->src_id, entry->src_data[0]);
1408 dev_err(adev->dev, " Can't decode VM fault info here on SRIOV VF\n");
1409 return 0;
1410 }
1411
1412 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1413 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1414 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1415 /* reset addr and status */
1416 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1417
1418 if (!addr && !status)
1419 return 0;
1420
1421 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1422 gmc_v8_0_set_fault_enable_default(adev, false);
1423
1424 if (printk_ratelimit()) {
1425 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1426 entry->src_id, entry->src_data[0]);
1427 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
1428 addr);
1429 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1430 status);
1431 gmc_v8_0_vm_decode_fault(adev, status, addr, mc_client,
1432 entry->pasid);
1433 }
1434
1435 return 0;
1436}
1437
1438static void fiji_update_mc_medium_grain_clock_gating(struct amdgpu_device *adev,
1439 bool enable)
1440{
1441 uint32_t data;
1442
1443 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) {
1444 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1445 data |= MC_HUB_MISC_HUB_CG__ENABLE_MASK;
1446 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1447
1448 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1449 data |= MC_HUB_MISC_SIP_CG__ENABLE_MASK;
1450 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1451
1452 data = RREG32(mmMC_HUB_MISC_VM_CG);
1453 data |= MC_HUB_MISC_VM_CG__ENABLE_MASK;
1454 WREG32(mmMC_HUB_MISC_VM_CG, data);
1455
1456 data = RREG32(mmMC_XPB_CLK_GAT);
1457 data |= MC_XPB_CLK_GAT__ENABLE_MASK;
1458 WREG32(mmMC_XPB_CLK_GAT, data);
1459
1460 data = RREG32(mmATC_MISC_CG);
1461 data |= ATC_MISC_CG__ENABLE_MASK;
1462 WREG32(mmATC_MISC_CG, data);
1463
1464 data = RREG32(mmMC_CITF_MISC_WR_CG);
1465 data |= MC_CITF_MISC_WR_CG__ENABLE_MASK;
1466 WREG32(mmMC_CITF_MISC_WR_CG, data);
1467
1468 data = RREG32(mmMC_CITF_MISC_RD_CG);
1469 data |= MC_CITF_MISC_RD_CG__ENABLE_MASK;
1470 WREG32(mmMC_CITF_MISC_RD_CG, data);
1471
1472 data = RREG32(mmMC_CITF_MISC_VM_CG);
1473 data |= MC_CITF_MISC_VM_CG__ENABLE_MASK;
1474 WREG32(mmMC_CITF_MISC_VM_CG, data);
1475
1476 data = RREG32(mmVM_L2_CG);
1477 data |= VM_L2_CG__ENABLE_MASK;
1478 WREG32(mmVM_L2_CG, data);
1479 } else {
1480 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1481 data &= ~MC_HUB_MISC_HUB_CG__ENABLE_MASK;
1482 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1483
1484 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1485 data &= ~MC_HUB_MISC_SIP_CG__ENABLE_MASK;
1486 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1487
1488 data = RREG32(mmMC_HUB_MISC_VM_CG);
1489 data &= ~MC_HUB_MISC_VM_CG__ENABLE_MASK;
1490 WREG32(mmMC_HUB_MISC_VM_CG, data);
1491
1492 data = RREG32(mmMC_XPB_CLK_GAT);
1493 data &= ~MC_XPB_CLK_GAT__ENABLE_MASK;
1494 WREG32(mmMC_XPB_CLK_GAT, data);
1495
1496 data = RREG32(mmATC_MISC_CG);
1497 data &= ~ATC_MISC_CG__ENABLE_MASK;
1498 WREG32(mmATC_MISC_CG, data);
1499
1500 data = RREG32(mmMC_CITF_MISC_WR_CG);
1501 data &= ~MC_CITF_MISC_WR_CG__ENABLE_MASK;
1502 WREG32(mmMC_CITF_MISC_WR_CG, data);
1503
1504 data = RREG32(mmMC_CITF_MISC_RD_CG);
1505 data &= ~MC_CITF_MISC_RD_CG__ENABLE_MASK;
1506 WREG32(mmMC_CITF_MISC_RD_CG, data);
1507
1508 data = RREG32(mmMC_CITF_MISC_VM_CG);
1509 data &= ~MC_CITF_MISC_VM_CG__ENABLE_MASK;
1510 WREG32(mmMC_CITF_MISC_VM_CG, data);
1511
1512 data = RREG32(mmVM_L2_CG);
1513 data &= ~VM_L2_CG__ENABLE_MASK;
1514 WREG32(mmVM_L2_CG, data);
1515 }
1516}
1517
1518static void fiji_update_mc_light_sleep(struct amdgpu_device *adev,
1519 bool enable)
1520{
1521 uint32_t data;
1522
1523 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) {
1524 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1525 data |= MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK;
1526 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1527
1528 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1529 data |= MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK;
1530 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1531
1532 data = RREG32(mmMC_HUB_MISC_VM_CG);
1533 data |= MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1534 WREG32(mmMC_HUB_MISC_VM_CG, data);
1535
1536 data = RREG32(mmMC_XPB_CLK_GAT);
1537 data |= MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK;
1538 WREG32(mmMC_XPB_CLK_GAT, data);
1539
1540 data = RREG32(mmATC_MISC_CG);
1541 data |= ATC_MISC_CG__MEM_LS_ENABLE_MASK;
1542 WREG32(mmATC_MISC_CG, data);
1543
1544 data = RREG32(mmMC_CITF_MISC_WR_CG);
1545 data |= MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK;
1546 WREG32(mmMC_CITF_MISC_WR_CG, data);
1547
1548 data = RREG32(mmMC_CITF_MISC_RD_CG);
1549 data |= MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK;
1550 WREG32(mmMC_CITF_MISC_RD_CG, data);
1551
1552 data = RREG32(mmMC_CITF_MISC_VM_CG);
1553 data |= MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1554 WREG32(mmMC_CITF_MISC_VM_CG, data);
1555
1556 data = RREG32(mmVM_L2_CG);
1557 data |= VM_L2_CG__MEM_LS_ENABLE_MASK;
1558 WREG32(mmVM_L2_CG, data);
1559 } else {
1560 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1561 data &= ~MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK;
1562 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1563
1564 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1565 data &= ~MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK;
1566 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1567
1568 data = RREG32(mmMC_HUB_MISC_VM_CG);
1569 data &= ~MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1570 WREG32(mmMC_HUB_MISC_VM_CG, data);
1571
1572 data = RREG32(mmMC_XPB_CLK_GAT);
1573 data &= ~MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK;
1574 WREG32(mmMC_XPB_CLK_GAT, data);
1575
1576 data = RREG32(mmATC_MISC_CG);
1577 data &= ~ATC_MISC_CG__MEM_LS_ENABLE_MASK;
1578 WREG32(mmATC_MISC_CG, data);
1579
1580 data = RREG32(mmMC_CITF_MISC_WR_CG);
1581 data &= ~MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK;
1582 WREG32(mmMC_CITF_MISC_WR_CG, data);
1583
1584 data = RREG32(mmMC_CITF_MISC_RD_CG);
1585 data &= ~MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK;
1586 WREG32(mmMC_CITF_MISC_RD_CG, data);
1587
1588 data = RREG32(mmMC_CITF_MISC_VM_CG);
1589 data &= ~MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1590 WREG32(mmMC_CITF_MISC_VM_CG, data);
1591
1592 data = RREG32(mmVM_L2_CG);
1593 data &= ~VM_L2_CG__MEM_LS_ENABLE_MASK;
1594 WREG32(mmVM_L2_CG, data);
1595 }
1596}
1597
1598static int gmc_v8_0_set_clockgating_state(void *handle,
1599 enum amd_clockgating_state state)
1600{
1601 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1602
1603 if (amdgpu_sriov_vf(adev))
1604 return 0;
1605
1606 switch (adev->asic_type) {
1607 case CHIP_FIJI:
1608 fiji_update_mc_medium_grain_clock_gating(adev,
1609 state == AMD_CG_STATE_GATE);
1610 fiji_update_mc_light_sleep(adev,
1611 state == AMD_CG_STATE_GATE);
1612 break;
1613 default:
1614 break;
1615 }
1616 return 0;
1617}
1618
1619static int gmc_v8_0_set_powergating_state(void *handle,
1620 enum amd_powergating_state state)
1621{
1622 return 0;
1623}
1624
1625static void gmc_v8_0_get_clockgating_state(void *handle, u32 *flags)
1626{
1627 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1628 int data;
1629
1630 if (amdgpu_sriov_vf(adev))
1631 *flags = 0;
1632
1633 /* AMD_CG_SUPPORT_MC_MGCG */
1634 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1635 if (data & MC_HUB_MISC_HUB_CG__ENABLE_MASK)
1636 *flags |= AMD_CG_SUPPORT_MC_MGCG;
1637
1638 /* AMD_CG_SUPPORT_MC_LS */
1639 if (data & MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK)
1640 *flags |= AMD_CG_SUPPORT_MC_LS;
1641}
1642
1643static const struct amd_ip_funcs gmc_v8_0_ip_funcs = {
1644 .name = "gmc_v8_0",
1645 .early_init = gmc_v8_0_early_init,
1646 .late_init = gmc_v8_0_late_init,
1647 .sw_init = gmc_v8_0_sw_init,
1648 .sw_fini = gmc_v8_0_sw_fini,
1649 .hw_init = gmc_v8_0_hw_init,
1650 .hw_fini = gmc_v8_0_hw_fini,
1651 .suspend = gmc_v8_0_suspend,
1652 .resume = gmc_v8_0_resume,
1653 .is_idle = gmc_v8_0_is_idle,
1654 .wait_for_idle = gmc_v8_0_wait_for_idle,
1655 .check_soft_reset = gmc_v8_0_check_soft_reset,
1656 .pre_soft_reset = gmc_v8_0_pre_soft_reset,
1657 .soft_reset = gmc_v8_0_soft_reset,
1658 .post_soft_reset = gmc_v8_0_post_soft_reset,
1659 .set_clockgating_state = gmc_v8_0_set_clockgating_state,
1660 .set_powergating_state = gmc_v8_0_set_powergating_state,
1661 .get_clockgating_state = gmc_v8_0_get_clockgating_state,
1662};
1663
1664static const struct amdgpu_gmc_funcs gmc_v8_0_gmc_funcs = {
1665 .flush_gpu_tlb = gmc_v8_0_flush_gpu_tlb,
1666 .emit_flush_gpu_tlb = gmc_v8_0_emit_flush_gpu_tlb,
1667 .emit_pasid_mapping = gmc_v8_0_emit_pasid_mapping,
1668 .set_pte_pde = gmc_v8_0_set_pte_pde,
1669 .set_prt = gmc_v8_0_set_prt,
1670 .get_vm_pte_flags = gmc_v8_0_get_vm_pte_flags,
1671 .get_vm_pde = gmc_v8_0_get_vm_pde
1672};
1673
1674static const struct amdgpu_irq_src_funcs gmc_v8_0_irq_funcs = {
1675 .set = gmc_v8_0_vm_fault_interrupt_state,
1676 .process = gmc_v8_0_process_interrupt,
1677};
1678
1679static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev)
1680{
1681 if (adev->gmc.gmc_funcs == NULL)
1682 adev->gmc.gmc_funcs = &gmc_v8_0_gmc_funcs;
1683}
1684
1685static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev)
1686{
1687 adev->gmc.vm_fault.num_types = 1;
1688 adev->gmc.vm_fault.funcs = &gmc_v8_0_irq_funcs;
1689}
1690
1691const struct amdgpu_ip_block_version gmc_v8_0_ip_block =
1692{
1693 .type = AMD_IP_BLOCK_TYPE_GMC,
1694 .major = 8,
1695 .minor = 0,
1696 .rev = 0,
1697 .funcs = &gmc_v8_0_ip_funcs,
1698};
1699
1700const struct amdgpu_ip_block_version gmc_v8_1_ip_block =
1701{
1702 .type = AMD_IP_BLOCK_TYPE_GMC,
1703 .major = 8,
1704 .minor = 1,
1705 .rev = 0,
1706 .funcs = &gmc_v8_0_ip_funcs,
1707};
1708
1709const struct amdgpu_ip_block_version gmc_v8_5_ip_block =
1710{
1711 .type = AMD_IP_BLOCK_TYPE_GMC,
1712 .major = 8,
1713 .minor = 5,
1714 .rev = 0,
1715 .funcs = &gmc_v8_0_ip_funcs,
1716};
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
24#include <linux/firmware.h>
25#include <linux/module.h>
26#include <linux/pci.h>
27
28#include <drm/drm_cache.h>
29#include "amdgpu.h"
30#include "gmc_v8_0.h"
31#include "amdgpu_ucode.h"
32#include "amdgpu_amdkfd.h"
33#include "amdgpu_gem.h"
34
35#include "gmc/gmc_8_1_d.h"
36#include "gmc/gmc_8_1_sh_mask.h"
37
38#include "bif/bif_5_0_d.h"
39#include "bif/bif_5_0_sh_mask.h"
40
41#include "oss/oss_3_0_d.h"
42#include "oss/oss_3_0_sh_mask.h"
43
44#include "dce/dce_10_0_d.h"
45#include "dce/dce_10_0_sh_mask.h"
46
47#include "vid.h"
48#include "vi.h"
49
50#include "amdgpu_atombios.h"
51
52#include "ivsrcid/ivsrcid_vislands30.h"
53
54static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev);
55static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev);
56static int gmc_v8_0_wait_for_idle(void *handle);
57
58MODULE_FIRMWARE("amdgpu/tonga_mc.bin");
59MODULE_FIRMWARE("amdgpu/polaris11_mc.bin");
60MODULE_FIRMWARE("amdgpu/polaris10_mc.bin");
61MODULE_FIRMWARE("amdgpu/polaris12_mc.bin");
62MODULE_FIRMWARE("amdgpu/polaris11_k_mc.bin");
63MODULE_FIRMWARE("amdgpu/polaris10_k_mc.bin");
64MODULE_FIRMWARE("amdgpu/polaris12_k_mc.bin");
65
66static const u32 golden_settings_tonga_a11[] =
67{
68 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
69 mmMC_HUB_RDREQ_DMIF_LIMIT, 0x0000007f, 0x00000028,
70 mmMC_HUB_WDP_UMC, 0x00007fb6, 0x00000991,
71 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
72 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
73 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
74 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
75};
76
77static const u32 tonga_mgcg_cgcg_init[] =
78{
79 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
80};
81
82static const u32 golden_settings_fiji_a10[] =
83{
84 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
85 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
86 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
87 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff,
88};
89
90static const u32 fiji_mgcg_cgcg_init[] =
91{
92 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
93};
94
95static const u32 golden_settings_polaris11_a11[] =
96{
97 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
98 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
99 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
100 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
101};
102
103static const u32 golden_settings_polaris10_a11[] =
104{
105 mmMC_ARB_WTM_GRPWT_RD, 0x00000003, 0x00000000,
106 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
107 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
108 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
109 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
110};
111
112static const u32 cz_mgcg_cgcg_init[] =
113{
114 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
115};
116
117static const u32 stoney_mgcg_cgcg_init[] =
118{
119 mmATC_MISC_CG, 0xffffffff, 0x000c0200,
120 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
121};
122
123static const u32 golden_settings_stoney_common[] =
124{
125 mmMC_HUB_RDREQ_UVD, MC_HUB_RDREQ_UVD__PRESCALE_MASK, 0x00000004,
126 mmMC_RD_GRP_OTH, MC_RD_GRP_OTH__UVD_MASK, 0x00600000
127};
128
129static void gmc_v8_0_init_golden_registers(struct amdgpu_device *adev)
130{
131 switch (adev->asic_type) {
132 case CHIP_FIJI:
133 amdgpu_device_program_register_sequence(adev,
134 fiji_mgcg_cgcg_init,
135 ARRAY_SIZE(fiji_mgcg_cgcg_init));
136 amdgpu_device_program_register_sequence(adev,
137 golden_settings_fiji_a10,
138 ARRAY_SIZE(golden_settings_fiji_a10));
139 break;
140 case CHIP_TONGA:
141 amdgpu_device_program_register_sequence(adev,
142 tonga_mgcg_cgcg_init,
143 ARRAY_SIZE(tonga_mgcg_cgcg_init));
144 amdgpu_device_program_register_sequence(adev,
145 golden_settings_tonga_a11,
146 ARRAY_SIZE(golden_settings_tonga_a11));
147 break;
148 case CHIP_POLARIS11:
149 case CHIP_POLARIS12:
150 case CHIP_VEGAM:
151 amdgpu_device_program_register_sequence(adev,
152 golden_settings_polaris11_a11,
153 ARRAY_SIZE(golden_settings_polaris11_a11));
154 break;
155 case CHIP_POLARIS10:
156 amdgpu_device_program_register_sequence(adev,
157 golden_settings_polaris10_a11,
158 ARRAY_SIZE(golden_settings_polaris10_a11));
159 break;
160 case CHIP_CARRIZO:
161 amdgpu_device_program_register_sequence(adev,
162 cz_mgcg_cgcg_init,
163 ARRAY_SIZE(cz_mgcg_cgcg_init));
164 break;
165 case CHIP_STONEY:
166 amdgpu_device_program_register_sequence(adev,
167 stoney_mgcg_cgcg_init,
168 ARRAY_SIZE(stoney_mgcg_cgcg_init));
169 amdgpu_device_program_register_sequence(adev,
170 golden_settings_stoney_common,
171 ARRAY_SIZE(golden_settings_stoney_common));
172 break;
173 default:
174 break;
175 }
176}
177
178static void gmc_v8_0_mc_stop(struct amdgpu_device *adev)
179{
180 u32 blackout;
181
182 gmc_v8_0_wait_for_idle(adev);
183
184 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
185 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) {
186 /* Block CPU access */
187 WREG32(mmBIF_FB_EN, 0);
188 /* blackout the MC */
189 blackout = REG_SET_FIELD(blackout,
190 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 1);
191 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout);
192 }
193 /* wait for the MC to settle */
194 udelay(100);
195}
196
197static void gmc_v8_0_mc_resume(struct amdgpu_device *adev)
198{
199 u32 tmp;
200
201 /* unblackout the MC */
202 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
203 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
204 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp);
205 /* allow CPU access */
206 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1);
207 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1);
208 WREG32(mmBIF_FB_EN, tmp);
209}
210
211/**
212 * gmc_v8_0_init_microcode - load ucode images from disk
213 *
214 * @adev: amdgpu_device pointer
215 *
216 * Use the firmware interface to load the ucode images into
217 * the driver (not loaded into hw).
218 * Returns 0 on success, error on failure.
219 */
220static int gmc_v8_0_init_microcode(struct amdgpu_device *adev)
221{
222 const char *chip_name;
223 char fw_name[30];
224 int err;
225
226 DRM_DEBUG("\n");
227
228 switch (adev->asic_type) {
229 case CHIP_TONGA:
230 chip_name = "tonga";
231 break;
232 case CHIP_POLARIS11:
233 if (((adev->pdev->device == 0x67ef) &&
234 ((adev->pdev->revision == 0xe0) ||
235 (adev->pdev->revision == 0xe5))) ||
236 ((adev->pdev->device == 0x67ff) &&
237 ((adev->pdev->revision == 0xcf) ||
238 (adev->pdev->revision == 0xef) ||
239 (adev->pdev->revision == 0xff))))
240 chip_name = "polaris11_k";
241 else if ((adev->pdev->device == 0x67ef) &&
242 (adev->pdev->revision == 0xe2))
243 chip_name = "polaris11_k";
244 else
245 chip_name = "polaris11";
246 break;
247 case CHIP_POLARIS10:
248 if ((adev->pdev->device == 0x67df) &&
249 ((adev->pdev->revision == 0xe1) ||
250 (adev->pdev->revision == 0xf7)))
251 chip_name = "polaris10_k";
252 else
253 chip_name = "polaris10";
254 break;
255 case CHIP_POLARIS12:
256 if (((adev->pdev->device == 0x6987) &&
257 ((adev->pdev->revision == 0xc0) ||
258 (adev->pdev->revision == 0xc3))) ||
259 ((adev->pdev->device == 0x6981) &&
260 ((adev->pdev->revision == 0x00) ||
261 (adev->pdev->revision == 0x01) ||
262 (adev->pdev->revision == 0x10))))
263 chip_name = "polaris12_k";
264 else
265 chip_name = "polaris12";
266 break;
267 case CHIP_FIJI:
268 case CHIP_CARRIZO:
269 case CHIP_STONEY:
270 case CHIP_VEGAM:
271 return 0;
272 default: BUG();
273 }
274
275 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name);
276 err = request_firmware(&adev->gmc.fw, fw_name, adev->dev);
277 if (err)
278 goto out;
279 err = amdgpu_ucode_validate(adev->gmc.fw);
280
281out:
282 if (err) {
283 pr_err("mc: Failed to load firmware \"%s\"\n", fw_name);
284 release_firmware(adev->gmc.fw);
285 adev->gmc.fw = NULL;
286 }
287 return err;
288}
289
290/**
291 * gmc_v8_0_tonga_mc_load_microcode - load tonga MC ucode into the hw
292 *
293 * @adev: amdgpu_device pointer
294 *
295 * Load the GDDR MC ucode into the hw (VI).
296 * Returns 0 on success, error on failure.
297 */
298static int gmc_v8_0_tonga_mc_load_microcode(struct amdgpu_device *adev)
299{
300 const struct mc_firmware_header_v1_0 *hdr;
301 const __le32 *fw_data = NULL;
302 const __le32 *io_mc_regs = NULL;
303 u32 running;
304 int i, ucode_size, regs_size;
305
306 /* Skip MC ucode loading on SR-IOV capable boards.
307 * vbios does this for us in asic_init in that case.
308 * Skip MC ucode loading on VF, because hypervisor will do that
309 * for this adaptor.
310 */
311 if (amdgpu_sriov_bios(adev))
312 return 0;
313
314 if (!adev->gmc.fw)
315 return -EINVAL;
316
317 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
318 amdgpu_ucode_print_mc_hdr(&hdr->header);
319
320 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
321 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
322 io_mc_regs = (const __le32 *)
323 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
324 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
325 fw_data = (const __le32 *)
326 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
327
328 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN);
329
330 if (running == 0) {
331 /* reset the engine and set to writable */
332 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
333 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
334
335 /* load mc io regs */
336 for (i = 0; i < regs_size; i++) {
337 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
338 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
339 }
340 /* load the MC ucode */
341 for (i = 0; i < ucode_size; i++)
342 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
343
344 /* put the engine back into the active state */
345 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
346 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
347 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
348
349 /* wait for training to complete */
350 for (i = 0; i < adev->usec_timeout; i++) {
351 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
352 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0))
353 break;
354 udelay(1);
355 }
356 for (i = 0; i < adev->usec_timeout; i++) {
357 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
358 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1))
359 break;
360 udelay(1);
361 }
362 }
363
364 return 0;
365}
366
367static int gmc_v8_0_polaris_mc_load_microcode(struct amdgpu_device *adev)
368{
369 const struct mc_firmware_header_v1_0 *hdr;
370 const __le32 *fw_data = NULL;
371 const __le32 *io_mc_regs = NULL;
372 u32 data;
373 int i, ucode_size, regs_size;
374
375 /* Skip MC ucode loading on SR-IOV capable boards.
376 * vbios does this for us in asic_init in that case.
377 * Skip MC ucode loading on VF, because hypervisor will do that
378 * for this adaptor.
379 */
380 if (amdgpu_sriov_bios(adev))
381 return 0;
382
383 if (!adev->gmc.fw)
384 return -EINVAL;
385
386 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
387 amdgpu_ucode_print_mc_hdr(&hdr->header);
388
389 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
390 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
391 io_mc_regs = (const __le32 *)
392 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
393 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
394 fw_data = (const __le32 *)
395 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
396
397 data = RREG32(mmMC_SEQ_MISC0);
398 data &= ~(0x40);
399 WREG32(mmMC_SEQ_MISC0, data);
400
401 /* load mc io regs */
402 for (i = 0; i < regs_size; i++) {
403 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
404 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
405 }
406
407 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
408 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
409
410 /* load the MC ucode */
411 for (i = 0; i < ucode_size; i++)
412 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
413
414 /* put the engine back into the active state */
415 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
416 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
417 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
418
419 /* wait for training to complete */
420 for (i = 0; i < adev->usec_timeout; i++) {
421 data = RREG32(mmMC_SEQ_MISC0);
422 if (data & 0x80)
423 break;
424 udelay(1);
425 }
426
427 return 0;
428}
429
430static void gmc_v8_0_vram_gtt_location(struct amdgpu_device *adev,
431 struct amdgpu_gmc *mc)
432{
433 u64 base = 0;
434
435 if (!amdgpu_sriov_vf(adev))
436 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF;
437 base <<= 24;
438
439 amdgpu_gmc_vram_location(adev, mc, base);
440 amdgpu_gmc_gart_location(adev, mc);
441}
442
443/**
444 * gmc_v8_0_mc_program - program the GPU memory controller
445 *
446 * @adev: amdgpu_device pointer
447 *
448 * Set the location of vram, gart, and AGP in the GPU's
449 * physical address space (VI).
450 */
451static void gmc_v8_0_mc_program(struct amdgpu_device *adev)
452{
453 u32 tmp;
454 int i, j;
455
456 /* Initialize HDP */
457 for (i = 0, j = 0; i < 32; i++, j += 0x6) {
458 WREG32((0xb05 + j), 0x00000000);
459 WREG32((0xb06 + j), 0x00000000);
460 WREG32((0xb07 + j), 0x00000000);
461 WREG32((0xb08 + j), 0x00000000);
462 WREG32((0xb09 + j), 0x00000000);
463 }
464 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0);
465
466 if (gmc_v8_0_wait_for_idle((void *)adev)) {
467 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
468 }
469 if (adev->mode_info.num_crtc) {
470 /* Lockout access through VGA aperture*/
471 tmp = RREG32(mmVGA_HDP_CONTROL);
472 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
473 WREG32(mmVGA_HDP_CONTROL, tmp);
474
475 /* disable VGA render */
476 tmp = RREG32(mmVGA_RENDER_CONTROL);
477 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
478 WREG32(mmVGA_RENDER_CONTROL, tmp);
479 }
480 /* Update configuration */
481 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR,
482 adev->gmc.vram_start >> 12);
483 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR,
484 adev->gmc.vram_end >> 12);
485 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
486 adev->vram_scratch.gpu_addr >> 12);
487
488 if (amdgpu_sriov_vf(adev)) {
489 tmp = ((adev->gmc.vram_end >> 24) & 0xFFFF) << 16;
490 tmp |= ((adev->gmc.vram_start >> 24) & 0xFFFF);
491 WREG32(mmMC_VM_FB_LOCATION, tmp);
492 /* XXX double check these! */
493 WREG32(mmHDP_NONSURFACE_BASE, (adev->gmc.vram_start >> 8));
494 WREG32(mmHDP_NONSURFACE_INFO, (2 << 7) | (1 << 30));
495 WREG32(mmHDP_NONSURFACE_SIZE, 0x3FFFFFFF);
496 }
497
498 WREG32(mmMC_VM_AGP_BASE, 0);
499 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF);
500 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF);
501 if (gmc_v8_0_wait_for_idle((void *)adev)) {
502 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
503 }
504
505 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
506
507 tmp = RREG32(mmHDP_MISC_CNTL);
508 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0);
509 WREG32(mmHDP_MISC_CNTL, tmp);
510
511 tmp = RREG32(mmHDP_HOST_PATH_CNTL);
512 WREG32(mmHDP_HOST_PATH_CNTL, tmp);
513}
514
515/**
516 * gmc_v8_0_mc_init - initialize the memory controller driver params
517 *
518 * @adev: amdgpu_device pointer
519 *
520 * Look up the amount of vram, vram width, and decide how to place
521 * vram and gart within the GPU's physical address space (VI).
522 * Returns 0 for success.
523 */
524static int gmc_v8_0_mc_init(struct amdgpu_device *adev)
525{
526 int r;
527
528 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev);
529 if (!adev->gmc.vram_width) {
530 u32 tmp;
531 int chansize, numchan;
532
533 /* Get VRAM informations */
534 tmp = RREG32(mmMC_ARB_RAMCFG);
535 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) {
536 chansize = 64;
537 } else {
538 chansize = 32;
539 }
540 tmp = RREG32(mmMC_SHARED_CHMAP);
541 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) {
542 case 0:
543 default:
544 numchan = 1;
545 break;
546 case 1:
547 numchan = 2;
548 break;
549 case 2:
550 numchan = 4;
551 break;
552 case 3:
553 numchan = 8;
554 break;
555 case 4:
556 numchan = 3;
557 break;
558 case 5:
559 numchan = 6;
560 break;
561 case 6:
562 numchan = 10;
563 break;
564 case 7:
565 numchan = 12;
566 break;
567 case 8:
568 numchan = 16;
569 break;
570 }
571 adev->gmc.vram_width = numchan * chansize;
572 }
573 /* size in MB on si */
574 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
575 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
576
577 if (!(adev->flags & AMD_IS_APU)) {
578 r = amdgpu_device_resize_fb_bar(adev);
579 if (r)
580 return r;
581 }
582 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
583 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
584
585#ifdef CONFIG_X86_64
586 if (adev->flags & AMD_IS_APU) {
587 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22;
588 adev->gmc.aper_size = adev->gmc.real_vram_size;
589 }
590#endif
591
592 /* In case the PCI BAR is larger than the actual amount of vram */
593 adev->gmc.visible_vram_size = adev->gmc.aper_size;
594 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
595 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
596
597 /* set the gart size */
598 if (amdgpu_gart_size == -1) {
599 switch (adev->asic_type) {
600 case CHIP_POLARIS10: /* all engines support GPUVM */
601 case CHIP_POLARIS11: /* all engines support GPUVM */
602 case CHIP_POLARIS12: /* all engines support GPUVM */
603 case CHIP_VEGAM: /* all engines support GPUVM */
604 default:
605 adev->gmc.gart_size = 256ULL << 20;
606 break;
607 case CHIP_TONGA: /* UVD, VCE do not support GPUVM */
608 case CHIP_FIJI: /* UVD, VCE do not support GPUVM */
609 case CHIP_CARRIZO: /* UVD, VCE do not support GPUVM, DCE SG support */
610 case CHIP_STONEY: /* UVD does not support GPUVM, DCE SG support */
611 adev->gmc.gart_size = 1024ULL << 20;
612 break;
613 }
614 } else {
615 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
616 }
617
618 gmc_v8_0_vram_gtt_location(adev, &adev->gmc);
619
620 return 0;
621}
622
623/*
624 * GART
625 * VMID 0 is the physical GPU addresses as used by the kernel.
626 * VMIDs 1-15 are used for userspace clients and are handled
627 * by the amdgpu vm/hsa code.
628 */
629
630/**
631 * gmc_v8_0_flush_gpu_tlb - gart tlb flush callback
632 *
633 * @adev: amdgpu_device pointer
634 * @vmid: vm instance to flush
635 *
636 * Flush the TLB for the requested page table (VI).
637 */
638static void gmc_v8_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
639 uint32_t vmhub, uint32_t flush_type)
640{
641 /* bits 0-15 are the VM contexts0-15 */
642 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
643}
644
645static uint64_t gmc_v8_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
646 unsigned vmid, uint64_t pd_addr)
647{
648 uint32_t reg;
649
650 if (vmid < 8)
651 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid;
652 else
653 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8;
654 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12);
655
656 /* bits 0-15 are the VM contexts0-15 */
657 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid);
658
659 return pd_addr;
660}
661
662static void gmc_v8_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
663 unsigned pasid)
664{
665 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid);
666}
667
668/*
669 * PTE format on VI:
670 * 63:40 reserved
671 * 39:12 4k physical page base address
672 * 11:7 fragment
673 * 6 write
674 * 5 read
675 * 4 exe
676 * 3 reserved
677 * 2 snooped
678 * 1 system
679 * 0 valid
680 *
681 * PDE format on VI:
682 * 63:59 block fragment size
683 * 58:40 reserved
684 * 39:1 physical base address of PTE
685 * bits 5:1 must be 0.
686 * 0 valid
687 */
688
689static uint64_t gmc_v8_0_get_vm_pte_flags(struct amdgpu_device *adev,
690 uint32_t flags)
691{
692 uint64_t pte_flag = 0;
693
694 if (flags & AMDGPU_VM_PAGE_EXECUTABLE)
695 pte_flag |= AMDGPU_PTE_EXECUTABLE;
696 if (flags & AMDGPU_VM_PAGE_READABLE)
697 pte_flag |= AMDGPU_PTE_READABLE;
698 if (flags & AMDGPU_VM_PAGE_WRITEABLE)
699 pte_flag |= AMDGPU_PTE_WRITEABLE;
700 if (flags & AMDGPU_VM_PAGE_PRT)
701 pte_flag |= AMDGPU_PTE_PRT;
702
703 return pte_flag;
704}
705
706static void gmc_v8_0_get_vm_pde(struct amdgpu_device *adev, int level,
707 uint64_t *addr, uint64_t *flags)
708{
709 BUG_ON(*addr & 0xFFFFFF0000000FFFULL);
710}
711
712/**
713 * gmc_v8_0_set_fault_enable_default - update VM fault handling
714 *
715 * @adev: amdgpu_device pointer
716 * @value: true redirects VM faults to the default page
717 */
718static void gmc_v8_0_set_fault_enable_default(struct amdgpu_device *adev,
719 bool value)
720{
721 u32 tmp;
722
723 tmp = RREG32(mmVM_CONTEXT1_CNTL);
724 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
725 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
726 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
727 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
728 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
729 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
730 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
731 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
732 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
733 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
734 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
735 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
736 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
737 EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
738 WREG32(mmVM_CONTEXT1_CNTL, tmp);
739}
740
741/**
742 * gmc_v8_0_set_prt - set PRT VM fault
743 *
744 * @adev: amdgpu_device pointer
745 * @enable: enable/disable VM fault handling for PRT
746*/
747static void gmc_v8_0_set_prt(struct amdgpu_device *adev, bool enable)
748{
749 u32 tmp;
750
751 if (enable && !adev->gmc.prt_warning) {
752 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n");
753 adev->gmc.prt_warning = true;
754 }
755
756 tmp = RREG32(mmVM_PRT_CNTL);
757 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
758 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
759 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
760 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
761 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
762 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
763 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
764 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
765 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
766 L2_CACHE_STORE_INVALID_ENTRIES, enable);
767 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
768 L1_TLB_STORE_INVALID_ENTRIES, enable);
769 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
770 MASK_PDE0_FAULT, enable);
771 WREG32(mmVM_PRT_CNTL, tmp);
772
773 if (enable) {
774 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT;
775 uint32_t high = adev->vm_manager.max_pfn -
776 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT);
777
778 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low);
779 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low);
780 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low);
781 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low);
782 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high);
783 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high);
784 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high);
785 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high);
786 } else {
787 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff);
788 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff);
789 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff);
790 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff);
791 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0);
792 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0);
793 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0);
794 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0);
795 }
796}
797
798/**
799 * gmc_v8_0_gart_enable - gart enable
800 *
801 * @adev: amdgpu_device pointer
802 *
803 * This sets up the TLBs, programs the page tables for VMID0,
804 * sets up the hw for VMIDs 1-15 which are allocated on
805 * demand, and sets up the global locations for the LDS, GDS,
806 * and GPUVM for FSA64 clients (VI).
807 * Returns 0 for success, errors for failure.
808 */
809static int gmc_v8_0_gart_enable(struct amdgpu_device *adev)
810{
811 uint64_t table_addr;
812 int r, i;
813 u32 tmp, field;
814
815 if (adev->gart.bo == NULL) {
816 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
817 return -EINVAL;
818 }
819 r = amdgpu_gart_table_vram_pin(adev);
820 if (r)
821 return r;
822
823 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
824
825 /* Setup TLB control */
826 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
827 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
828 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
829 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
830 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
831 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
832 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
833 /* Setup L2 cache */
834 tmp = RREG32(mmVM_L2_CNTL);
835 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
836 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
837 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
838 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
839 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
840 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
841 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1);
842 WREG32(mmVM_L2_CNTL, tmp);
843 tmp = RREG32(mmVM_L2_CNTL2);
844 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
845 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
846 WREG32(mmVM_L2_CNTL2, tmp);
847
848 field = adev->vm_manager.fragment_size;
849 tmp = RREG32(mmVM_L2_CNTL3);
850 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
851 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field);
852 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field);
853 WREG32(mmVM_L2_CNTL3, tmp);
854 /* XXX: set to enable PTE/PDE in system memory */
855 tmp = RREG32(mmVM_L2_CNTL4);
856 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_PHYSICAL, 0);
857 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SHARED, 0);
858 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PDE_REQUEST_SNOOP, 0);
859 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_PHYSICAL, 0);
860 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SHARED, 0);
861 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT0_PTE_REQUEST_SNOOP, 0);
862 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_PHYSICAL, 0);
863 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SHARED, 0);
864 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PDE_REQUEST_SNOOP, 0);
865 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_PHYSICAL, 0);
866 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SHARED, 0);
867 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL4, VMC_TAP_CONTEXT1_PTE_REQUEST_SNOOP, 0);
868 WREG32(mmVM_L2_CNTL4, tmp);
869 /* setup context0 */
870 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12);
871 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12);
872 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12);
873 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
874 (u32)(adev->dummy_page_addr >> 12));
875 WREG32(mmVM_CONTEXT0_CNTL2, 0);
876 tmp = RREG32(mmVM_CONTEXT0_CNTL);
877 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
878 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
879 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
880 WREG32(mmVM_CONTEXT0_CNTL, tmp);
881
882 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_LOW_ADDR, 0);
883 WREG32(mmVM_L2_CONTEXT1_IDENTITY_APERTURE_HIGH_ADDR, 0);
884 WREG32(mmVM_L2_CONTEXT_IDENTITY_PHYSICAL_OFFSET, 0);
885
886 /* empty context1-15 */
887 /* FIXME start with 4G, once using 2 level pt switch to full
888 * vm size space
889 */
890 /* set vm size, must be a multiple of 4 */
891 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
892 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
893 for (i = 1; i < 16; i++) {
894 if (i < 8)
895 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
896 table_addr >> 12);
897 else
898 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
899 table_addr >> 12);
900 }
901
902 /* enable context1-15 */
903 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
904 (u32)(adev->dummy_page_addr >> 12));
905 WREG32(mmVM_CONTEXT1_CNTL2, 4);
906 tmp = RREG32(mmVM_CONTEXT1_CNTL);
907 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
908 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
909 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
910 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
911 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
912 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, VALID_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
913 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, READ_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
914 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
915 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, EXECUTE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
916 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
917 adev->vm_manager.block_size - 9);
918 WREG32(mmVM_CONTEXT1_CNTL, tmp);
919 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
920 gmc_v8_0_set_fault_enable_default(adev, false);
921 else
922 gmc_v8_0_set_fault_enable_default(adev, true);
923
924 gmc_v8_0_flush_gpu_tlb(adev, 0, 0, 0);
925 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
926 (unsigned)(adev->gmc.gart_size >> 20),
927 (unsigned long long)table_addr);
928 adev->gart.ready = true;
929 return 0;
930}
931
932static int gmc_v8_0_gart_init(struct amdgpu_device *adev)
933{
934 int r;
935
936 if (adev->gart.bo) {
937 WARN(1, "R600 PCIE GART already initialized\n");
938 return 0;
939 }
940 /* Initialize common gart structure */
941 r = amdgpu_gart_init(adev);
942 if (r)
943 return r;
944 adev->gart.table_size = adev->gart.num_gpu_pages * 8;
945 adev->gart.gart_pte_flags = AMDGPU_PTE_EXECUTABLE;
946 return amdgpu_gart_table_vram_alloc(adev);
947}
948
949/**
950 * gmc_v8_0_gart_disable - gart disable
951 *
952 * @adev: amdgpu_device pointer
953 *
954 * This disables all VM page table (VI).
955 */
956static void gmc_v8_0_gart_disable(struct amdgpu_device *adev)
957{
958 u32 tmp;
959
960 /* Disable all tables */
961 WREG32(mmVM_CONTEXT0_CNTL, 0);
962 WREG32(mmVM_CONTEXT1_CNTL, 0);
963 /* Setup TLB control */
964 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
965 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
966 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
967 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
968 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
969 /* Setup L2 cache */
970 tmp = RREG32(mmVM_L2_CNTL);
971 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
972 WREG32(mmVM_L2_CNTL, tmp);
973 WREG32(mmVM_L2_CNTL2, 0);
974 amdgpu_gart_table_vram_unpin(adev);
975}
976
977/**
978 * gmc_v8_0_vm_decode_fault - print human readable fault info
979 *
980 * @adev: amdgpu_device pointer
981 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
982 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
983 *
984 * Print human readable fault information (VI).
985 */
986static void gmc_v8_0_vm_decode_fault(struct amdgpu_device *adev, u32 status,
987 u32 addr, u32 mc_client, unsigned pasid)
988{
989 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
990 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
991 PROTECTIONS);
992 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
993 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
994 u32 mc_id;
995
996 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
997 MEMORY_CLIENT_ID);
998
999 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
1000 protections, vmid, pasid, addr,
1001 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1002 MEMORY_CLIENT_RW) ?
1003 "write" : "read", block, mc_client, mc_id);
1004}
1005
1006static int gmc_v8_0_convert_vram_type(int mc_seq_vram_type)
1007{
1008 switch (mc_seq_vram_type) {
1009 case MC_SEQ_MISC0__MT__GDDR1:
1010 return AMDGPU_VRAM_TYPE_GDDR1;
1011 case MC_SEQ_MISC0__MT__DDR2:
1012 return AMDGPU_VRAM_TYPE_DDR2;
1013 case MC_SEQ_MISC0__MT__GDDR3:
1014 return AMDGPU_VRAM_TYPE_GDDR3;
1015 case MC_SEQ_MISC0__MT__GDDR4:
1016 return AMDGPU_VRAM_TYPE_GDDR4;
1017 case MC_SEQ_MISC0__MT__GDDR5:
1018 return AMDGPU_VRAM_TYPE_GDDR5;
1019 case MC_SEQ_MISC0__MT__HBM:
1020 return AMDGPU_VRAM_TYPE_HBM;
1021 case MC_SEQ_MISC0__MT__DDR3:
1022 return AMDGPU_VRAM_TYPE_DDR3;
1023 default:
1024 return AMDGPU_VRAM_TYPE_UNKNOWN;
1025 }
1026}
1027
1028static int gmc_v8_0_early_init(void *handle)
1029{
1030 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1031
1032 gmc_v8_0_set_gmc_funcs(adev);
1033 gmc_v8_0_set_irq_funcs(adev);
1034
1035 adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
1036 adev->gmc.shared_aperture_end =
1037 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
1038 adev->gmc.private_aperture_start =
1039 adev->gmc.shared_aperture_end + 1;
1040 adev->gmc.private_aperture_end =
1041 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
1042
1043 return 0;
1044}
1045
1046static int gmc_v8_0_late_init(void *handle)
1047{
1048 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1049
1050 amdgpu_bo_late_init(adev);
1051
1052 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS)
1053 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
1054 else
1055 return 0;
1056}
1057
1058static unsigned gmc_v8_0_get_vbios_fb_size(struct amdgpu_device *adev)
1059{
1060 u32 d1vga_control = RREG32(mmD1VGA_CONTROL);
1061 unsigned size;
1062
1063 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
1064 size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
1065 } else {
1066 u32 viewport = RREG32(mmVIEWPORT_SIZE);
1067 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
1068 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) *
1069 4);
1070 }
1071 /* return 0 if the pre-OS buffer uses up most of vram */
1072 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1073 return 0;
1074 return size;
1075}
1076
1077#define mmMC_SEQ_MISC0_FIJI 0xA71
1078
1079static int gmc_v8_0_sw_init(void *handle)
1080{
1081 int r;
1082 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1083
1084 adev->num_vmhubs = 1;
1085
1086 if (adev->flags & AMD_IS_APU) {
1087 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
1088 } else {
1089 u32 tmp;
1090
1091 if ((adev->asic_type == CHIP_FIJI) ||
1092 (adev->asic_type == CHIP_VEGAM))
1093 tmp = RREG32(mmMC_SEQ_MISC0_FIJI);
1094 else
1095 tmp = RREG32(mmMC_SEQ_MISC0);
1096 tmp &= MC_SEQ_MISC0__MT__MASK;
1097 adev->gmc.vram_type = gmc_v8_0_convert_vram_type(tmp);
1098 }
1099
1100 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault);
1101 if (r)
1102 return r;
1103
1104 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault);
1105 if (r)
1106 return r;
1107
1108 /* Adjust VM size here.
1109 * Currently set to 4GB ((1 << 20) 4k pages).
1110 * Max GPUVM size for cayman and SI is 40 bits.
1111 */
1112 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40);
1113
1114 /* Set the internal MC address mask
1115 * This is the max address of the GPU's
1116 * internal address space.
1117 */
1118 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
1119
1120 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40));
1121 if (r) {
1122 pr_warn("amdgpu: No suitable DMA available\n");
1123 return r;
1124 }
1125 adev->need_swiotlb = drm_need_swiotlb(40);
1126
1127 r = gmc_v8_0_init_microcode(adev);
1128 if (r) {
1129 DRM_ERROR("Failed to load mc firmware!\n");
1130 return r;
1131 }
1132
1133 r = gmc_v8_0_mc_init(adev);
1134 if (r)
1135 return r;
1136
1137 adev->gmc.stolen_size = gmc_v8_0_get_vbios_fb_size(adev);
1138
1139 /* Memory manager */
1140 r = amdgpu_bo_init(adev);
1141 if (r)
1142 return r;
1143
1144 r = gmc_v8_0_gart_init(adev);
1145 if (r)
1146 return r;
1147
1148 /*
1149 * number of VMs
1150 * VMID 0 is reserved for System
1151 * amdgpu graphics/compute will use VMIDs 1-7
1152 * amdkfd will use VMIDs 8-15
1153 */
1154 adev->vm_manager.id_mgr[0].num_ids = AMDGPU_NUM_OF_VMIDS;
1155 amdgpu_vm_manager_init(adev);
1156
1157 /* base offset of vram pages */
1158 if (adev->flags & AMD_IS_APU) {
1159 u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
1160
1161 tmp <<= 22;
1162 adev->vm_manager.vram_base_offset = tmp;
1163 } else {
1164 adev->vm_manager.vram_base_offset = 0;
1165 }
1166
1167 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info),
1168 GFP_KERNEL);
1169 if (!adev->gmc.vm_fault_info)
1170 return -ENOMEM;
1171 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1172
1173 return 0;
1174}
1175
1176static int gmc_v8_0_sw_fini(void *handle)
1177{
1178 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1179
1180 amdgpu_gem_force_release(adev);
1181 amdgpu_vm_manager_fini(adev);
1182 kfree(adev->gmc.vm_fault_info);
1183 amdgpu_gart_table_vram_free(adev);
1184 amdgpu_bo_fini(adev);
1185 amdgpu_gart_fini(adev);
1186 release_firmware(adev->gmc.fw);
1187 adev->gmc.fw = NULL;
1188
1189 return 0;
1190}
1191
1192static int gmc_v8_0_hw_init(void *handle)
1193{
1194 int r;
1195 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1196
1197 gmc_v8_0_init_golden_registers(adev);
1198
1199 gmc_v8_0_mc_program(adev);
1200
1201 if (adev->asic_type == CHIP_TONGA) {
1202 r = gmc_v8_0_tonga_mc_load_microcode(adev);
1203 if (r) {
1204 DRM_ERROR("Failed to load MC firmware!\n");
1205 return r;
1206 }
1207 } else if (adev->asic_type == CHIP_POLARIS11 ||
1208 adev->asic_type == CHIP_POLARIS10 ||
1209 adev->asic_type == CHIP_POLARIS12) {
1210 r = gmc_v8_0_polaris_mc_load_microcode(adev);
1211 if (r) {
1212 DRM_ERROR("Failed to load MC firmware!\n");
1213 return r;
1214 }
1215 }
1216
1217 r = gmc_v8_0_gart_enable(adev);
1218 if (r)
1219 return r;
1220
1221 return r;
1222}
1223
1224static int gmc_v8_0_hw_fini(void *handle)
1225{
1226 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1227
1228 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1229 gmc_v8_0_gart_disable(adev);
1230
1231 return 0;
1232}
1233
1234static int gmc_v8_0_suspend(void *handle)
1235{
1236 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1237
1238 gmc_v8_0_hw_fini(adev);
1239
1240 return 0;
1241}
1242
1243static int gmc_v8_0_resume(void *handle)
1244{
1245 int r;
1246 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1247
1248 r = gmc_v8_0_hw_init(adev);
1249 if (r)
1250 return r;
1251
1252 amdgpu_vmid_reset_all(adev);
1253
1254 return 0;
1255}
1256
1257static bool gmc_v8_0_is_idle(void *handle)
1258{
1259 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1260 u32 tmp = RREG32(mmSRBM_STATUS);
1261
1262 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1263 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1264 return false;
1265
1266 return true;
1267}
1268
1269static int gmc_v8_0_wait_for_idle(void *handle)
1270{
1271 unsigned i;
1272 u32 tmp;
1273 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1274
1275 for (i = 0; i < adev->usec_timeout; i++) {
1276 /* read MC_STATUS */
1277 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1278 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1279 SRBM_STATUS__MCC_BUSY_MASK |
1280 SRBM_STATUS__MCD_BUSY_MASK |
1281 SRBM_STATUS__VMC_BUSY_MASK |
1282 SRBM_STATUS__VMC1_BUSY_MASK);
1283 if (!tmp)
1284 return 0;
1285 udelay(1);
1286 }
1287 return -ETIMEDOUT;
1288
1289}
1290
1291static bool gmc_v8_0_check_soft_reset(void *handle)
1292{
1293 u32 srbm_soft_reset = 0;
1294 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1295 u32 tmp = RREG32(mmSRBM_STATUS);
1296
1297 if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1298 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1299 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1300
1301 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1302 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1303 if (!(adev->flags & AMD_IS_APU))
1304 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1305 SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1306 }
1307 if (srbm_soft_reset) {
1308 adev->gmc.srbm_soft_reset = srbm_soft_reset;
1309 return true;
1310 } else {
1311 adev->gmc.srbm_soft_reset = 0;
1312 return false;
1313 }
1314}
1315
1316static int gmc_v8_0_pre_soft_reset(void *handle)
1317{
1318 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1319
1320 if (!adev->gmc.srbm_soft_reset)
1321 return 0;
1322
1323 gmc_v8_0_mc_stop(adev);
1324 if (gmc_v8_0_wait_for_idle(adev)) {
1325 dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1326 }
1327
1328 return 0;
1329}
1330
1331static int gmc_v8_0_soft_reset(void *handle)
1332{
1333 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1334 u32 srbm_soft_reset;
1335
1336 if (!adev->gmc.srbm_soft_reset)
1337 return 0;
1338 srbm_soft_reset = adev->gmc.srbm_soft_reset;
1339
1340 if (srbm_soft_reset) {
1341 u32 tmp;
1342
1343 tmp = RREG32(mmSRBM_SOFT_RESET);
1344 tmp |= srbm_soft_reset;
1345 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1346 WREG32(mmSRBM_SOFT_RESET, tmp);
1347 tmp = RREG32(mmSRBM_SOFT_RESET);
1348
1349 udelay(50);
1350
1351 tmp &= ~srbm_soft_reset;
1352 WREG32(mmSRBM_SOFT_RESET, tmp);
1353 tmp = RREG32(mmSRBM_SOFT_RESET);
1354
1355 /* Wait a little for things to settle down */
1356 udelay(50);
1357 }
1358
1359 return 0;
1360}
1361
1362static int gmc_v8_0_post_soft_reset(void *handle)
1363{
1364 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1365
1366 if (!adev->gmc.srbm_soft_reset)
1367 return 0;
1368
1369 gmc_v8_0_mc_resume(adev);
1370 return 0;
1371}
1372
1373static int gmc_v8_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1374 struct amdgpu_irq_src *src,
1375 unsigned type,
1376 enum amdgpu_interrupt_state state)
1377{
1378 u32 tmp;
1379 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1380 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1381 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1382 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1383 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1384 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1385 VM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1386
1387 switch (state) {
1388 case AMDGPU_IRQ_STATE_DISABLE:
1389 /* system context */
1390 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1391 tmp &= ~bits;
1392 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1393 /* VMs */
1394 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1395 tmp &= ~bits;
1396 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1397 break;
1398 case AMDGPU_IRQ_STATE_ENABLE:
1399 /* system context */
1400 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1401 tmp |= bits;
1402 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1403 /* VMs */
1404 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1405 tmp |= bits;
1406 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1407 break;
1408 default:
1409 break;
1410 }
1411
1412 return 0;
1413}
1414
1415static int gmc_v8_0_process_interrupt(struct amdgpu_device *adev,
1416 struct amdgpu_irq_src *source,
1417 struct amdgpu_iv_entry *entry)
1418{
1419 u32 addr, status, mc_client, vmid;
1420
1421 if (amdgpu_sriov_vf(adev)) {
1422 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1423 entry->src_id, entry->src_data[0]);
1424 dev_err(adev->dev, " Can't decode VM fault info here on SRIOV VF\n");
1425 return 0;
1426 }
1427
1428 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1429 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1430 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1431 /* reset addr and status */
1432 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1433
1434 if (!addr && !status)
1435 return 0;
1436
1437 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1438 gmc_v8_0_set_fault_enable_default(adev, false);
1439
1440 if (printk_ratelimit()) {
1441 struct amdgpu_task_info task_info;
1442
1443 memset(&task_info, 0, sizeof(struct amdgpu_task_info));
1444 amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
1445
1446 dev_err(adev->dev, "GPU fault detected: %d 0x%08x for process %s pid %d thread %s pid %d\n",
1447 entry->src_id, entry->src_data[0], task_info.process_name,
1448 task_info.tgid, task_info.task_name, task_info.pid);
1449 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
1450 addr);
1451 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1452 status);
1453 gmc_v8_0_vm_decode_fault(adev, status, addr, mc_client,
1454 entry->pasid);
1455 }
1456
1457 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1458 VMID);
1459 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid)
1460 && !atomic_read(&adev->gmc.vm_fault_info_updated)) {
1461 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info;
1462 u32 protections = REG_GET_FIELD(status,
1463 VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1464 PROTECTIONS);
1465
1466 info->vmid = vmid;
1467 info->mc_id = REG_GET_FIELD(status,
1468 VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1469 MEMORY_CLIENT_ID);
1470 info->status = status;
1471 info->page_addr = addr;
1472 info->prot_valid = protections & 0x7 ? true : false;
1473 info->prot_read = protections & 0x8 ? true : false;
1474 info->prot_write = protections & 0x10 ? true : false;
1475 info->prot_exec = protections & 0x20 ? true : false;
1476 mb();
1477 atomic_set(&adev->gmc.vm_fault_info_updated, 1);
1478 }
1479
1480 return 0;
1481}
1482
1483static void fiji_update_mc_medium_grain_clock_gating(struct amdgpu_device *adev,
1484 bool enable)
1485{
1486 uint32_t data;
1487
1488 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG)) {
1489 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1490 data |= MC_HUB_MISC_HUB_CG__ENABLE_MASK;
1491 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1492
1493 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1494 data |= MC_HUB_MISC_SIP_CG__ENABLE_MASK;
1495 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1496
1497 data = RREG32(mmMC_HUB_MISC_VM_CG);
1498 data |= MC_HUB_MISC_VM_CG__ENABLE_MASK;
1499 WREG32(mmMC_HUB_MISC_VM_CG, data);
1500
1501 data = RREG32(mmMC_XPB_CLK_GAT);
1502 data |= MC_XPB_CLK_GAT__ENABLE_MASK;
1503 WREG32(mmMC_XPB_CLK_GAT, data);
1504
1505 data = RREG32(mmATC_MISC_CG);
1506 data |= ATC_MISC_CG__ENABLE_MASK;
1507 WREG32(mmATC_MISC_CG, data);
1508
1509 data = RREG32(mmMC_CITF_MISC_WR_CG);
1510 data |= MC_CITF_MISC_WR_CG__ENABLE_MASK;
1511 WREG32(mmMC_CITF_MISC_WR_CG, data);
1512
1513 data = RREG32(mmMC_CITF_MISC_RD_CG);
1514 data |= MC_CITF_MISC_RD_CG__ENABLE_MASK;
1515 WREG32(mmMC_CITF_MISC_RD_CG, data);
1516
1517 data = RREG32(mmMC_CITF_MISC_VM_CG);
1518 data |= MC_CITF_MISC_VM_CG__ENABLE_MASK;
1519 WREG32(mmMC_CITF_MISC_VM_CG, data);
1520
1521 data = RREG32(mmVM_L2_CG);
1522 data |= VM_L2_CG__ENABLE_MASK;
1523 WREG32(mmVM_L2_CG, data);
1524 } else {
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 }
1561}
1562
1563static void fiji_update_mc_light_sleep(struct amdgpu_device *adev,
1564 bool enable)
1565{
1566 uint32_t data;
1567
1568 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS)) {
1569 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1570 data |= MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK;
1571 WREG32(mmMC_HUB_MISC_HUB_CG, data);
1572
1573 data = RREG32(mmMC_HUB_MISC_SIP_CG);
1574 data |= MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK;
1575 WREG32(mmMC_HUB_MISC_SIP_CG, data);
1576
1577 data = RREG32(mmMC_HUB_MISC_VM_CG);
1578 data |= MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1579 WREG32(mmMC_HUB_MISC_VM_CG, data);
1580
1581 data = RREG32(mmMC_XPB_CLK_GAT);
1582 data |= MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK;
1583 WREG32(mmMC_XPB_CLK_GAT, data);
1584
1585 data = RREG32(mmATC_MISC_CG);
1586 data |= ATC_MISC_CG__MEM_LS_ENABLE_MASK;
1587 WREG32(mmATC_MISC_CG, data);
1588
1589 data = RREG32(mmMC_CITF_MISC_WR_CG);
1590 data |= MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK;
1591 WREG32(mmMC_CITF_MISC_WR_CG, data);
1592
1593 data = RREG32(mmMC_CITF_MISC_RD_CG);
1594 data |= MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK;
1595 WREG32(mmMC_CITF_MISC_RD_CG, data);
1596
1597 data = RREG32(mmMC_CITF_MISC_VM_CG);
1598 data |= MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK;
1599 WREG32(mmMC_CITF_MISC_VM_CG, data);
1600
1601 data = RREG32(mmVM_L2_CG);
1602 data |= VM_L2_CG__MEM_LS_ENABLE_MASK;
1603 WREG32(mmVM_L2_CG, data);
1604 } else {
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 }
1641}
1642
1643static int gmc_v8_0_set_clockgating_state(void *handle,
1644 enum amd_clockgating_state state)
1645{
1646 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1647
1648 if (amdgpu_sriov_vf(adev))
1649 return 0;
1650
1651 switch (adev->asic_type) {
1652 case CHIP_FIJI:
1653 fiji_update_mc_medium_grain_clock_gating(adev,
1654 state == AMD_CG_STATE_GATE);
1655 fiji_update_mc_light_sleep(adev,
1656 state == AMD_CG_STATE_GATE);
1657 break;
1658 default:
1659 break;
1660 }
1661 return 0;
1662}
1663
1664static int gmc_v8_0_set_powergating_state(void *handle,
1665 enum amd_powergating_state state)
1666{
1667 return 0;
1668}
1669
1670static void gmc_v8_0_get_clockgating_state(void *handle, u32 *flags)
1671{
1672 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1673 int data;
1674
1675 if (amdgpu_sriov_vf(adev))
1676 *flags = 0;
1677
1678 /* AMD_CG_SUPPORT_MC_MGCG */
1679 data = RREG32(mmMC_HUB_MISC_HUB_CG);
1680 if (data & MC_HUB_MISC_HUB_CG__ENABLE_MASK)
1681 *flags |= AMD_CG_SUPPORT_MC_MGCG;
1682
1683 /* AMD_CG_SUPPORT_MC_LS */
1684 if (data & MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK)
1685 *flags |= AMD_CG_SUPPORT_MC_LS;
1686}
1687
1688static const struct amd_ip_funcs gmc_v8_0_ip_funcs = {
1689 .name = "gmc_v8_0",
1690 .early_init = gmc_v8_0_early_init,
1691 .late_init = gmc_v8_0_late_init,
1692 .sw_init = gmc_v8_0_sw_init,
1693 .sw_fini = gmc_v8_0_sw_fini,
1694 .hw_init = gmc_v8_0_hw_init,
1695 .hw_fini = gmc_v8_0_hw_fini,
1696 .suspend = gmc_v8_0_suspend,
1697 .resume = gmc_v8_0_resume,
1698 .is_idle = gmc_v8_0_is_idle,
1699 .wait_for_idle = gmc_v8_0_wait_for_idle,
1700 .check_soft_reset = gmc_v8_0_check_soft_reset,
1701 .pre_soft_reset = gmc_v8_0_pre_soft_reset,
1702 .soft_reset = gmc_v8_0_soft_reset,
1703 .post_soft_reset = gmc_v8_0_post_soft_reset,
1704 .set_clockgating_state = gmc_v8_0_set_clockgating_state,
1705 .set_powergating_state = gmc_v8_0_set_powergating_state,
1706 .get_clockgating_state = gmc_v8_0_get_clockgating_state,
1707};
1708
1709static const struct amdgpu_gmc_funcs gmc_v8_0_gmc_funcs = {
1710 .flush_gpu_tlb = gmc_v8_0_flush_gpu_tlb,
1711 .emit_flush_gpu_tlb = gmc_v8_0_emit_flush_gpu_tlb,
1712 .emit_pasid_mapping = gmc_v8_0_emit_pasid_mapping,
1713 .set_prt = gmc_v8_0_set_prt,
1714 .get_vm_pte_flags = gmc_v8_0_get_vm_pte_flags,
1715 .get_vm_pde = gmc_v8_0_get_vm_pde
1716};
1717
1718static const struct amdgpu_irq_src_funcs gmc_v8_0_irq_funcs = {
1719 .set = gmc_v8_0_vm_fault_interrupt_state,
1720 .process = gmc_v8_0_process_interrupt,
1721};
1722
1723static void gmc_v8_0_set_gmc_funcs(struct amdgpu_device *adev)
1724{
1725 adev->gmc.gmc_funcs = &gmc_v8_0_gmc_funcs;
1726}
1727
1728static void gmc_v8_0_set_irq_funcs(struct amdgpu_device *adev)
1729{
1730 adev->gmc.vm_fault.num_types = 1;
1731 adev->gmc.vm_fault.funcs = &gmc_v8_0_irq_funcs;
1732}
1733
1734const struct amdgpu_ip_block_version gmc_v8_0_ip_block =
1735{
1736 .type = AMD_IP_BLOCK_TYPE_GMC,
1737 .major = 8,
1738 .minor = 0,
1739 .rev = 0,
1740 .funcs = &gmc_v8_0_ip_funcs,
1741};
1742
1743const struct amdgpu_ip_block_version gmc_v8_1_ip_block =
1744{
1745 .type = AMD_IP_BLOCK_TYPE_GMC,
1746 .major = 8,
1747 .minor = 1,
1748 .rev = 0,
1749 .funcs = &gmc_v8_0_ip_funcs,
1750};
1751
1752const struct amdgpu_ip_block_version gmc_v8_5_ip_block =
1753{
1754 .type = AMD_IP_BLOCK_TYPE_GMC,
1755 .major = 8,
1756 .minor = 5,
1757 .rev = 0,
1758 .funcs = &gmc_v8_0_ip_funcs,
1759};