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 "cikd.h"
31 #include "cik.h"
32 #include "gmc_v7_0.h"
33 #include "amdgpu_ucode.h"
34 #include "amdgpu_amdkfd.h"
35 #include "amdgpu_gem.h"
36
37 #include "bif/bif_4_1_d.h"
38 #include "bif/bif_4_1_sh_mask.h"
39
40 #include "gmc/gmc_7_1_d.h"
41 #include "gmc/gmc_7_1_sh_mask.h"
42
43 #include "oss/oss_2_0_d.h"
44 #include "oss/oss_2_0_sh_mask.h"
45
46 #include "dce/dce_8_0_d.h"
47 #include "dce/dce_8_0_sh_mask.h"
48
49 #include "amdgpu_atombios.h"
50
51 #include "ivsrcid/ivsrcid_vislands30.h"
52
53 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev);
54 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev);
55 static int gmc_v7_0_wait_for_idle(void *handle);
56
57 MODULE_FIRMWARE("amdgpu/bonaire_mc.bin");
58 MODULE_FIRMWARE("amdgpu/hawaii_mc.bin");
59 MODULE_FIRMWARE("amdgpu/topaz_mc.bin");
60
61 static const u32 golden_settings_iceland_a11[] =
62 {
63 mmVM_PRT_APERTURE0_LOW_ADDR, 0x0fffffff, 0x0fffffff,
64 mmVM_PRT_APERTURE1_LOW_ADDR, 0x0fffffff, 0x0fffffff,
65 mmVM_PRT_APERTURE2_LOW_ADDR, 0x0fffffff, 0x0fffffff,
66 mmVM_PRT_APERTURE3_LOW_ADDR, 0x0fffffff, 0x0fffffff
67 };
68
69 static const u32 iceland_mgcg_cgcg_init[] =
70 {
71 mmMC_MEM_POWER_LS, 0xffffffff, 0x00000104
72 };
73
gmc_v7_0_init_golden_registers(struct amdgpu_device * adev)74 static void gmc_v7_0_init_golden_registers(struct amdgpu_device *adev)
75 {
76 switch (adev->asic_type) {
77 case CHIP_TOPAZ:
78 amdgpu_device_program_register_sequence(adev,
79 iceland_mgcg_cgcg_init,
80 ARRAY_SIZE(iceland_mgcg_cgcg_init));
81 amdgpu_device_program_register_sequence(adev,
82 golden_settings_iceland_a11,
83 ARRAY_SIZE(golden_settings_iceland_a11));
84 break;
85 default:
86 break;
87 }
88 }
89
gmc_v7_0_mc_stop(struct amdgpu_device * adev)90 static void gmc_v7_0_mc_stop(struct amdgpu_device *adev)
91 {
92 u32 blackout;
93
94 gmc_v7_0_wait_for_idle((void *)adev);
95
96 blackout = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
97 if (REG_GET_FIELD(blackout, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE) != 1) {
98 /* Block CPU access */
99 WREG32(mmBIF_FB_EN, 0);
100 /* blackout the MC */
101 blackout = REG_SET_FIELD(blackout,
102 MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
103 WREG32(mmMC_SHARED_BLACKOUT_CNTL, blackout | 1);
104 }
105 /* wait for the MC to settle */
106 udelay(100);
107 }
108
gmc_v7_0_mc_resume(struct amdgpu_device * adev)109 static void gmc_v7_0_mc_resume(struct amdgpu_device *adev)
110 {
111 u32 tmp;
112
113 /* unblackout the MC */
114 tmp = RREG32(mmMC_SHARED_BLACKOUT_CNTL);
115 tmp = REG_SET_FIELD(tmp, MC_SHARED_BLACKOUT_CNTL, BLACKOUT_MODE, 0);
116 WREG32(mmMC_SHARED_BLACKOUT_CNTL, tmp);
117 /* allow CPU access */
118 tmp = REG_SET_FIELD(0, BIF_FB_EN, FB_READ_EN, 1);
119 tmp = REG_SET_FIELD(tmp, BIF_FB_EN, FB_WRITE_EN, 1);
120 WREG32(mmBIF_FB_EN, tmp);
121 }
122
123 /**
124 * gmc_v7_0_init_microcode - load ucode images from disk
125 *
126 * @adev: amdgpu_device pointer
127 *
128 * Use the firmware interface to load the ucode images into
129 * the driver (not loaded into hw).
130 * Returns 0 on success, error on failure.
131 */
gmc_v7_0_init_microcode(struct amdgpu_device * adev)132 static int gmc_v7_0_init_microcode(struct amdgpu_device *adev)
133 {
134 const char *chip_name;
135 char fw_name[30];
136 int err;
137
138 DRM_DEBUG("\n");
139
140 switch (adev->asic_type) {
141 case CHIP_BONAIRE:
142 chip_name = "bonaire";
143 break;
144 case CHIP_HAWAII:
145 chip_name = "hawaii";
146 break;
147 case CHIP_TOPAZ:
148 chip_name = "topaz";
149 break;
150 case CHIP_KAVERI:
151 case CHIP_KABINI:
152 case CHIP_MULLINS:
153 return 0;
154 default: BUG();
155 }
156
157 snprintf(fw_name, sizeof(fw_name), "amdgpu/%s_mc.bin", chip_name);
158
159 err = request_firmware(&adev->gmc.fw, fw_name, adev->dev);
160 if (err)
161 goto out;
162 err = amdgpu_ucode_validate(adev->gmc.fw);
163
164 out:
165 if (err) {
166 pr_err("cik_mc: Failed to load firmware \"%s\"\n", fw_name);
167 release_firmware(adev->gmc.fw);
168 adev->gmc.fw = NULL;
169 }
170 return err;
171 }
172
173 /**
174 * gmc_v7_0_mc_load_microcode - load MC ucode into the hw
175 *
176 * @adev: amdgpu_device pointer
177 *
178 * Load the GDDR MC ucode into the hw (CIK).
179 * Returns 0 on success, error on failure.
180 */
gmc_v7_0_mc_load_microcode(struct amdgpu_device * adev)181 static int gmc_v7_0_mc_load_microcode(struct amdgpu_device *adev)
182 {
183 const struct mc_firmware_header_v1_0 *hdr;
184 const __le32 *fw_data = NULL;
185 const __le32 *io_mc_regs = NULL;
186 u32 running;
187 int i, ucode_size, regs_size;
188
189 if (!adev->gmc.fw)
190 return -EINVAL;
191
192 hdr = (const struct mc_firmware_header_v1_0 *)adev->gmc.fw->data;
193 amdgpu_ucode_print_mc_hdr(&hdr->header);
194
195 adev->gmc.fw_version = le32_to_cpu(hdr->header.ucode_version);
196 regs_size = le32_to_cpu(hdr->io_debug_size_bytes) / (4 * 2);
197 io_mc_regs = (const __le32 *)
198 (adev->gmc.fw->data + le32_to_cpu(hdr->io_debug_array_offset_bytes));
199 ucode_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4;
200 fw_data = (const __le32 *)
201 (adev->gmc.fw->data + le32_to_cpu(hdr->header.ucode_array_offset_bytes));
202
203 running = REG_GET_FIELD(RREG32(mmMC_SEQ_SUP_CNTL), MC_SEQ_SUP_CNTL, RUN);
204
205 if (running == 0) {
206 /* reset the engine and set to writable */
207 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
208 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000010);
209
210 /* load mc io regs */
211 for (i = 0; i < regs_size; i++) {
212 WREG32(mmMC_SEQ_IO_DEBUG_INDEX, le32_to_cpup(io_mc_regs++));
213 WREG32(mmMC_SEQ_IO_DEBUG_DATA, le32_to_cpup(io_mc_regs++));
214 }
215 /* load the MC ucode */
216 for (i = 0; i < ucode_size; i++)
217 WREG32(mmMC_SEQ_SUP_PGM, le32_to_cpup(fw_data++));
218
219 /* put the engine back into the active state */
220 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000008);
221 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000004);
222 WREG32(mmMC_SEQ_SUP_CNTL, 0x00000001);
223
224 /* wait for training to complete */
225 for (i = 0; i < adev->usec_timeout; i++) {
226 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
227 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D0))
228 break;
229 udelay(1);
230 }
231 for (i = 0; i < adev->usec_timeout; i++) {
232 if (REG_GET_FIELD(RREG32(mmMC_SEQ_TRAIN_WAKEUP_CNTL),
233 MC_SEQ_TRAIN_WAKEUP_CNTL, TRAIN_DONE_D1))
234 break;
235 udelay(1);
236 }
237 }
238
239 return 0;
240 }
241
gmc_v7_0_vram_gtt_location(struct amdgpu_device * adev,struct amdgpu_gmc * mc)242 static void gmc_v7_0_vram_gtt_location(struct amdgpu_device *adev,
243 struct amdgpu_gmc *mc)
244 {
245 u64 base = RREG32(mmMC_VM_FB_LOCATION) & 0xFFFF;
246 base <<= 24;
247
248 amdgpu_gmc_vram_location(adev, mc, base);
249 amdgpu_gmc_gart_location(adev, mc);
250 }
251
252 /**
253 * gmc_v7_0_mc_program - program the GPU memory controller
254 *
255 * @adev: amdgpu_device pointer
256 *
257 * Set the location of vram, gart, and AGP in the GPU's
258 * physical address space (CIK).
259 */
gmc_v7_0_mc_program(struct amdgpu_device * adev)260 static void gmc_v7_0_mc_program(struct amdgpu_device *adev)
261 {
262 u32 tmp;
263 int i, j;
264
265 /* Initialize HDP */
266 for (i = 0, j = 0; i < 32; i++, j += 0x6) {
267 WREG32((0xb05 + j), 0x00000000);
268 WREG32((0xb06 + j), 0x00000000);
269 WREG32((0xb07 + j), 0x00000000);
270 WREG32((0xb08 + j), 0x00000000);
271 WREG32((0xb09 + j), 0x00000000);
272 }
273 WREG32(mmHDP_REG_COHERENCY_FLUSH_CNTL, 0);
274
275 if (gmc_v7_0_wait_for_idle((void *)adev)) {
276 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
277 }
278 if (adev->mode_info.num_crtc) {
279 /* Lockout access through VGA aperture*/
280 tmp = RREG32(mmVGA_HDP_CONTROL);
281 tmp = REG_SET_FIELD(tmp, VGA_HDP_CONTROL, VGA_MEMORY_DISABLE, 1);
282 WREG32(mmVGA_HDP_CONTROL, tmp);
283
284 /* disable VGA render */
285 tmp = RREG32(mmVGA_RENDER_CONTROL);
286 tmp = REG_SET_FIELD(tmp, VGA_RENDER_CONTROL, VGA_VSTATUS_CNTL, 0);
287 WREG32(mmVGA_RENDER_CONTROL, tmp);
288 }
289 /* Update configuration */
290 WREG32(mmMC_VM_SYSTEM_APERTURE_LOW_ADDR,
291 adev->gmc.vram_start >> 12);
292 WREG32(mmMC_VM_SYSTEM_APERTURE_HIGH_ADDR,
293 adev->gmc.vram_end >> 12);
294 WREG32(mmMC_VM_SYSTEM_APERTURE_DEFAULT_ADDR,
295 adev->vram_scratch.gpu_addr >> 12);
296 WREG32(mmMC_VM_AGP_BASE, 0);
297 WREG32(mmMC_VM_AGP_TOP, 0x0FFFFFFF);
298 WREG32(mmMC_VM_AGP_BOT, 0x0FFFFFFF);
299 if (gmc_v7_0_wait_for_idle((void *)adev)) {
300 dev_warn(adev->dev, "Wait for MC idle timedout !\n");
301 }
302
303 WREG32(mmBIF_FB_EN, BIF_FB_EN__FB_READ_EN_MASK | BIF_FB_EN__FB_WRITE_EN_MASK);
304
305 tmp = RREG32(mmHDP_MISC_CNTL);
306 tmp = REG_SET_FIELD(tmp, HDP_MISC_CNTL, FLUSH_INVALIDATE_CACHE, 0);
307 WREG32(mmHDP_MISC_CNTL, tmp);
308
309 tmp = RREG32(mmHDP_HOST_PATH_CNTL);
310 WREG32(mmHDP_HOST_PATH_CNTL, tmp);
311 }
312
313 /**
314 * gmc_v7_0_mc_init - initialize the memory controller driver params
315 *
316 * @adev: amdgpu_device pointer
317 *
318 * Look up the amount of vram, vram width, and decide how to place
319 * vram and gart within the GPU's physical address space (CIK).
320 * Returns 0 for success.
321 */
gmc_v7_0_mc_init(struct amdgpu_device * adev)322 static int gmc_v7_0_mc_init(struct amdgpu_device *adev)
323 {
324 int r;
325
326 adev->gmc.vram_width = amdgpu_atombios_get_vram_width(adev);
327 if (!adev->gmc.vram_width) {
328 u32 tmp;
329 int chansize, numchan;
330
331 /* Get VRAM informations */
332 tmp = RREG32(mmMC_ARB_RAMCFG);
333 if (REG_GET_FIELD(tmp, MC_ARB_RAMCFG, CHANSIZE)) {
334 chansize = 64;
335 } else {
336 chansize = 32;
337 }
338 tmp = RREG32(mmMC_SHARED_CHMAP);
339 switch (REG_GET_FIELD(tmp, MC_SHARED_CHMAP, NOOFCHAN)) {
340 case 0:
341 default:
342 numchan = 1;
343 break;
344 case 1:
345 numchan = 2;
346 break;
347 case 2:
348 numchan = 4;
349 break;
350 case 3:
351 numchan = 8;
352 break;
353 case 4:
354 numchan = 3;
355 break;
356 case 5:
357 numchan = 6;
358 break;
359 case 6:
360 numchan = 10;
361 break;
362 case 7:
363 numchan = 12;
364 break;
365 case 8:
366 numchan = 16;
367 break;
368 }
369 adev->gmc.vram_width = numchan * chansize;
370 }
371 /* size in MB on si */
372 adev->gmc.mc_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
373 adev->gmc.real_vram_size = RREG32(mmCONFIG_MEMSIZE) * 1024ULL * 1024ULL;
374
375 if (!(adev->flags & AMD_IS_APU)) {
376 r = amdgpu_device_resize_fb_bar(adev);
377 if (r)
378 return r;
379 }
380 adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
381 adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
382
383 #ifdef CONFIG_X86_64
384 if ((adev->flags & AMD_IS_APU) &&
385 adev->gmc.real_vram_size > adev->gmc.aper_size &&
386 !amdgpu_passthrough(adev)) {
387 adev->gmc.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22;
388 adev->gmc.aper_size = adev->gmc.real_vram_size;
389 }
390 #endif
391
392 /* In case the PCI BAR is larger than the actual amount of vram */
393 adev->gmc.visible_vram_size = adev->gmc.aper_size;
394 if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
395 adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
396
397 /* set the gart size */
398 if (amdgpu_gart_size == -1) {
399 switch (adev->asic_type) {
400 case CHIP_TOPAZ: /* no MM engines */
401 default:
402 adev->gmc.gart_size = 256ULL << 20;
403 break;
404 #ifdef CONFIG_DRM_AMDGPU_CIK
405 case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */
406 case CHIP_HAWAII: /* UVD, VCE do not support GPUVM */
407 case CHIP_KAVERI: /* UVD, VCE do not support GPUVM */
408 case CHIP_KABINI: /* UVD, VCE do not support GPUVM */
409 case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */
410 adev->gmc.gart_size = 1024ULL << 20;
411 break;
412 #endif
413 }
414 } else {
415 adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
416 }
417
418 adev->gmc.gart_size += adev->pm.smu_prv_buffer_size;
419 gmc_v7_0_vram_gtt_location(adev, &adev->gmc);
420
421 return 0;
422 }
423
424 /**
425 * gmc_v7_0_flush_gpu_tlb_pasid - tlb flush via pasid
426 *
427 * @adev: amdgpu_device pointer
428 * @pasid: pasid to be flush
429 * @flush_type: type of flush
430 * @all_hub: flush all hubs
431 *
432 * Flush the TLB for the requested pasid.
433 */
gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device * adev,uint16_t pasid,uint32_t flush_type,bool all_hub)434 static int gmc_v7_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
435 uint16_t pasid, uint32_t flush_type,
436 bool all_hub)
437 {
438 int vmid;
439 unsigned int tmp;
440
441 if (amdgpu_in_reset(adev))
442 return -EIO;
443
444 for (vmid = 1; vmid < 16; vmid++) {
445
446 tmp = RREG32(mmATC_VMID0_PASID_MAPPING + vmid);
447 if ((tmp & ATC_VMID0_PASID_MAPPING__VALID_MASK) &&
448 (tmp & ATC_VMID0_PASID_MAPPING__PASID_MASK) == pasid) {
449 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
450 RREG32(mmVM_INVALIDATE_RESPONSE);
451 break;
452 }
453 }
454
455 return 0;
456 }
457
458 /*
459 * GART
460 * VMID 0 is the physical GPU addresses as used by the kernel.
461 * VMIDs 1-15 are used for userspace clients and are handled
462 * by the amdgpu vm/hsa code.
463 */
464
465 /**
466 * gmc_v7_0_flush_gpu_tlb - gart tlb flush callback
467 *
468 * @adev: amdgpu_device pointer
469 * @vmid: vm instance to flush
470 * @vmhub: which hub to flush
471 * @flush_type: type of flush
472 * *
473 * Flush the TLB for the requested page table (CIK).
474 */
gmc_v7_0_flush_gpu_tlb(struct amdgpu_device * adev,uint32_t vmid,uint32_t vmhub,uint32_t flush_type)475 static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
476 uint32_t vmhub, uint32_t flush_type)
477 {
478 /* bits 0-15 are the VM contexts0-15 */
479 WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
480 }
481
gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring * ring,unsigned vmid,uint64_t pd_addr)482 static uint64_t gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
483 unsigned vmid, uint64_t pd_addr)
484 {
485 uint32_t reg;
486
487 if (vmid < 8)
488 reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid;
489 else
490 reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8;
491 amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12);
492
493 /* bits 0-15 are the VM contexts0-15 */
494 amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid);
495
496 return pd_addr;
497 }
498
gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring * ring,unsigned vmid,unsigned pasid)499 static void gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
500 unsigned pasid)
501 {
502 amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid);
503 }
504
gmc_v7_0_get_vm_pde(struct amdgpu_device * adev,int level,uint64_t * addr,uint64_t * flags)505 static void gmc_v7_0_get_vm_pde(struct amdgpu_device *adev, int level,
506 uint64_t *addr, uint64_t *flags)
507 {
508 BUG_ON(*addr & 0xFFFFFF0000000FFFULL);
509 }
510
gmc_v7_0_get_vm_pte(struct amdgpu_device * adev,struct amdgpu_bo_va_mapping * mapping,uint64_t * flags)511 static void gmc_v7_0_get_vm_pte(struct amdgpu_device *adev,
512 struct amdgpu_bo_va_mapping *mapping,
513 uint64_t *flags)
514 {
515 *flags &= ~AMDGPU_PTE_EXECUTABLE;
516 *flags &= ~AMDGPU_PTE_PRT;
517 }
518
519 /**
520 * gmc_v7_0_set_fault_enable_default - update VM fault handling
521 *
522 * @adev: amdgpu_device pointer
523 * @value: true redirects VM faults to the default page
524 */
gmc_v7_0_set_fault_enable_default(struct amdgpu_device * adev,bool value)525 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev,
526 bool value)
527 {
528 u32 tmp;
529
530 tmp = RREG32(mmVM_CONTEXT1_CNTL);
531 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
532 RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
533 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
534 DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
535 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
536 PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
537 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
538 VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
539 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
540 READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
541 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
542 WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
543 WREG32(mmVM_CONTEXT1_CNTL, tmp);
544 }
545
546 /**
547 * gmc_v7_0_set_prt - set PRT VM fault
548 *
549 * @adev: amdgpu_device pointer
550 * @enable: enable/disable VM fault handling for PRT
551 */
gmc_v7_0_set_prt(struct amdgpu_device * adev,bool enable)552 static void gmc_v7_0_set_prt(struct amdgpu_device *adev, bool enable)
553 {
554 uint32_t tmp;
555
556 if (enable && !adev->gmc.prt_warning) {
557 dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n");
558 adev->gmc.prt_warning = true;
559 }
560
561 tmp = RREG32(mmVM_PRT_CNTL);
562 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
563 CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
564 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
565 CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
566 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
567 TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
568 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
569 TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
570 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
571 L2_CACHE_STORE_INVALID_ENTRIES, enable);
572 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
573 L1_TLB_STORE_INVALID_ENTRIES, enable);
574 tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
575 MASK_PDE0_FAULT, enable);
576 WREG32(mmVM_PRT_CNTL, tmp);
577
578 if (enable) {
579 uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT;
580 uint32_t high = adev->vm_manager.max_pfn -
581 (AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT);
582
583 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low);
584 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low);
585 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low);
586 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low);
587 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high);
588 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high);
589 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high);
590 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high);
591 } else {
592 WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff);
593 WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff);
594 WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff);
595 WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff);
596 WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0);
597 WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0);
598 WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0);
599 WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0);
600 }
601 }
602
603 /**
604 * gmc_v7_0_gart_enable - gart enable
605 *
606 * @adev: amdgpu_device pointer
607 *
608 * This sets up the TLBs, programs the page tables for VMID0,
609 * sets up the hw for VMIDs 1-15 which are allocated on
610 * demand, and sets up the global locations for the LDS, GDS,
611 * and GPUVM for FSA64 clients (CIK).
612 * Returns 0 for success, errors for failure.
613 */
gmc_v7_0_gart_enable(struct amdgpu_device * adev)614 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev)
615 {
616 uint64_t table_addr;
617 int r, i;
618 u32 tmp, field;
619
620 if (adev->gart.bo == NULL) {
621 dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
622 return -EINVAL;
623 }
624 r = amdgpu_gart_table_vram_pin(adev);
625 if (r)
626 return r;
627
628 table_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
629
630 /* Setup TLB control */
631 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
632 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
633 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
634 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
635 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
636 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
637 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
638 /* Setup L2 cache */
639 tmp = RREG32(mmVM_L2_CNTL);
640 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
641 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
642 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
643 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
644 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
645 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
646 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1);
647 WREG32(mmVM_L2_CNTL, tmp);
648 tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
649 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
650 WREG32(mmVM_L2_CNTL2, tmp);
651
652 field = adev->vm_manager.fragment_size;
653 tmp = RREG32(mmVM_L2_CNTL3);
654 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
655 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field);
656 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field);
657 WREG32(mmVM_L2_CNTL3, tmp);
658 /* setup context0 */
659 WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12);
660 WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12);
661 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12);
662 WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
663 (u32)(adev->dummy_page_addr >> 12));
664 WREG32(mmVM_CONTEXT0_CNTL2, 0);
665 tmp = RREG32(mmVM_CONTEXT0_CNTL);
666 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
667 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
668 tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
669 WREG32(mmVM_CONTEXT0_CNTL, tmp);
670
671 WREG32(0x575, 0);
672 WREG32(0x576, 0);
673 WREG32(0x577, 0);
674
675 /* empty context1-15 */
676 /* FIXME start with 4G, once using 2 level pt switch to full
677 * vm size space
678 */
679 /* set vm size, must be a multiple of 4 */
680 WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
681 WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
682 for (i = 1; i < AMDGPU_NUM_VMID; i++) {
683 if (i < 8)
684 WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
685 table_addr >> 12);
686 else
687 WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
688 table_addr >> 12);
689 }
690
691 /* enable context1-15 */
692 WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
693 (u32)(adev->dummy_page_addr >> 12));
694 WREG32(mmVM_CONTEXT1_CNTL2, 4);
695 tmp = RREG32(mmVM_CONTEXT1_CNTL);
696 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
697 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
698 tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
699 adev->vm_manager.block_size - 9);
700 WREG32(mmVM_CONTEXT1_CNTL, tmp);
701 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
702 gmc_v7_0_set_fault_enable_default(adev, false);
703 else
704 gmc_v7_0_set_fault_enable_default(adev, true);
705
706 if (adev->asic_type == CHIP_KAVERI) {
707 tmp = RREG32(mmCHUB_CONTROL);
708 tmp &= ~BYPASS_VM;
709 WREG32(mmCHUB_CONTROL, tmp);
710 }
711
712 gmc_v7_0_flush_gpu_tlb(adev, 0, 0, 0);
713 DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
714 (unsigned)(adev->gmc.gart_size >> 20),
715 (unsigned long long)table_addr);
716 adev->gart.ready = true;
717 return 0;
718 }
719
gmc_v7_0_gart_init(struct amdgpu_device * adev)720 static int gmc_v7_0_gart_init(struct amdgpu_device *adev)
721 {
722 int r;
723
724 if (adev->gart.bo) {
725 WARN(1, "R600 PCIE GART already initialized\n");
726 return 0;
727 }
728 /* Initialize common gart structure */
729 r = amdgpu_gart_init(adev);
730 if (r)
731 return r;
732 adev->gart.table_size = adev->gart.num_gpu_pages * 8;
733 adev->gart.gart_pte_flags = 0;
734 return amdgpu_gart_table_vram_alloc(adev);
735 }
736
737 /**
738 * gmc_v7_0_gart_disable - gart disable
739 *
740 * @adev: amdgpu_device pointer
741 *
742 * This disables all VM page table (CIK).
743 */
gmc_v7_0_gart_disable(struct amdgpu_device * adev)744 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev)
745 {
746 u32 tmp;
747
748 /* Disable all tables */
749 WREG32(mmVM_CONTEXT0_CNTL, 0);
750 WREG32(mmVM_CONTEXT1_CNTL, 0);
751 /* Setup TLB control */
752 tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
753 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
754 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
755 tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
756 WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
757 /* Setup L2 cache */
758 tmp = RREG32(mmVM_L2_CNTL);
759 tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
760 WREG32(mmVM_L2_CNTL, tmp);
761 WREG32(mmVM_L2_CNTL2, 0);
762 amdgpu_gart_table_vram_unpin(adev);
763 }
764
765 /**
766 * gmc_v7_0_vm_decode_fault - print human readable fault info
767 *
768 * @adev: amdgpu_device pointer
769 * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
770 * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
771 * @mc_client: VM_CONTEXT1_PROTECTION_FAULT_MCCLIENT register value
772 * @pasid: debug logging only - no functional use
773 *
774 * Print human readable fault information (CIK).
775 */
gmc_v7_0_vm_decode_fault(struct amdgpu_device * adev,u32 status,u32 addr,u32 mc_client,unsigned pasid)776 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev, u32 status,
777 u32 addr, u32 mc_client, unsigned pasid)
778 {
779 u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
780 u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
781 PROTECTIONS);
782 char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
783 (mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
784 u32 mc_id;
785
786 mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
787 MEMORY_CLIENT_ID);
788
789 dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
790 protections, vmid, pasid, addr,
791 REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
792 MEMORY_CLIENT_RW) ?
793 "write" : "read", block, mc_client, mc_id);
794 }
795
796
797 static const u32 mc_cg_registers[] = {
798 mmMC_HUB_MISC_HUB_CG,
799 mmMC_HUB_MISC_SIP_CG,
800 mmMC_HUB_MISC_VM_CG,
801 mmMC_XPB_CLK_GAT,
802 mmATC_MISC_CG,
803 mmMC_CITF_MISC_WR_CG,
804 mmMC_CITF_MISC_RD_CG,
805 mmMC_CITF_MISC_VM_CG,
806 mmVM_L2_CG,
807 };
808
809 static const u32 mc_cg_ls_en[] = {
810 MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK,
811 MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK,
812 MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK,
813 MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK,
814 ATC_MISC_CG__MEM_LS_ENABLE_MASK,
815 MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK,
816 MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK,
817 MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK,
818 VM_L2_CG__MEM_LS_ENABLE_MASK,
819 };
820
821 static const u32 mc_cg_en[] = {
822 MC_HUB_MISC_HUB_CG__ENABLE_MASK,
823 MC_HUB_MISC_SIP_CG__ENABLE_MASK,
824 MC_HUB_MISC_VM_CG__ENABLE_MASK,
825 MC_XPB_CLK_GAT__ENABLE_MASK,
826 ATC_MISC_CG__ENABLE_MASK,
827 MC_CITF_MISC_WR_CG__ENABLE_MASK,
828 MC_CITF_MISC_RD_CG__ENABLE_MASK,
829 MC_CITF_MISC_VM_CG__ENABLE_MASK,
830 VM_L2_CG__ENABLE_MASK,
831 };
832
gmc_v7_0_enable_mc_ls(struct amdgpu_device * adev,bool enable)833 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev,
834 bool enable)
835 {
836 int i;
837 u32 orig, data;
838
839 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
840 orig = data = RREG32(mc_cg_registers[i]);
841 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS))
842 data |= mc_cg_ls_en[i];
843 else
844 data &= ~mc_cg_ls_en[i];
845 if (data != orig)
846 WREG32(mc_cg_registers[i], data);
847 }
848 }
849
gmc_v7_0_enable_mc_mgcg(struct amdgpu_device * adev,bool enable)850 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev,
851 bool enable)
852 {
853 int i;
854 u32 orig, data;
855
856 for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
857 orig = data = RREG32(mc_cg_registers[i]);
858 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG))
859 data |= mc_cg_en[i];
860 else
861 data &= ~mc_cg_en[i];
862 if (data != orig)
863 WREG32(mc_cg_registers[i], data);
864 }
865 }
866
gmc_v7_0_enable_bif_mgls(struct amdgpu_device * adev,bool enable)867 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev,
868 bool enable)
869 {
870 u32 orig, data;
871
872 orig = data = RREG32_PCIE(ixPCIE_CNTL2);
873
874 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_BIF_LS)) {
875 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1);
876 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1);
877 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1);
878 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1);
879 } else {
880 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0);
881 data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0);
882 data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0);
883 data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0);
884 }
885
886 if (orig != data)
887 WREG32_PCIE(ixPCIE_CNTL2, data);
888 }
889
gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device * adev,bool enable)890 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev,
891 bool enable)
892 {
893 u32 orig, data;
894
895 orig = data = RREG32(mmHDP_HOST_PATH_CNTL);
896
897 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_MGCG))
898 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0);
899 else
900 data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1);
901
902 if (orig != data)
903 WREG32(mmHDP_HOST_PATH_CNTL, data);
904 }
905
gmc_v7_0_enable_hdp_ls(struct amdgpu_device * adev,bool enable)906 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev,
907 bool enable)
908 {
909 u32 orig, data;
910
911 orig = data = RREG32(mmHDP_MEM_POWER_LS);
912
913 if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_LS))
914 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1);
915 else
916 data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0);
917
918 if (orig != data)
919 WREG32(mmHDP_MEM_POWER_LS, data);
920 }
921
gmc_v7_0_convert_vram_type(int mc_seq_vram_type)922 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type)
923 {
924 switch (mc_seq_vram_type) {
925 case MC_SEQ_MISC0__MT__GDDR1:
926 return AMDGPU_VRAM_TYPE_GDDR1;
927 case MC_SEQ_MISC0__MT__DDR2:
928 return AMDGPU_VRAM_TYPE_DDR2;
929 case MC_SEQ_MISC0__MT__GDDR3:
930 return AMDGPU_VRAM_TYPE_GDDR3;
931 case MC_SEQ_MISC0__MT__GDDR4:
932 return AMDGPU_VRAM_TYPE_GDDR4;
933 case MC_SEQ_MISC0__MT__GDDR5:
934 return AMDGPU_VRAM_TYPE_GDDR5;
935 case MC_SEQ_MISC0__MT__HBM:
936 return AMDGPU_VRAM_TYPE_HBM;
937 case MC_SEQ_MISC0__MT__DDR3:
938 return AMDGPU_VRAM_TYPE_DDR3;
939 default:
940 return AMDGPU_VRAM_TYPE_UNKNOWN;
941 }
942 }
943
gmc_v7_0_early_init(void * handle)944 static int gmc_v7_0_early_init(void *handle)
945 {
946 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
947
948 gmc_v7_0_set_gmc_funcs(adev);
949 gmc_v7_0_set_irq_funcs(adev);
950
951 adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
952 adev->gmc.shared_aperture_end =
953 adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
954 adev->gmc.private_aperture_start =
955 adev->gmc.shared_aperture_end + 1;
956 adev->gmc.private_aperture_end =
957 adev->gmc.private_aperture_start + (4ULL << 30) - 1;
958
959 return 0;
960 }
961
gmc_v7_0_late_init(void * handle)962 static int gmc_v7_0_late_init(void *handle)
963 {
964 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
965
966 if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS)
967 return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
968 else
969 return 0;
970 }
971
gmc_v7_0_get_vbios_fb_size(struct amdgpu_device * adev)972 static unsigned gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev)
973 {
974 u32 d1vga_control = RREG32(mmD1VGA_CONTROL);
975 unsigned size;
976
977 if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
978 size = AMDGPU_VBIOS_VGA_ALLOCATION;
979 } else {
980 u32 viewport = RREG32(mmVIEWPORT_SIZE);
981 size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
982 REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) *
983 4);
984 }
985
986 return size;
987 }
988
gmc_v7_0_sw_init(void * handle)989 static int gmc_v7_0_sw_init(void *handle)
990 {
991 int r;
992 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
993
994 adev->num_vmhubs = 1;
995
996 if (adev->flags & AMD_IS_APU) {
997 adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
998 } else {
999 u32 tmp = RREG32(mmMC_SEQ_MISC0);
1000 tmp &= MC_SEQ_MISC0__MT__MASK;
1001 adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp);
1002 }
1003
1004 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault);
1005 if (r)
1006 return r;
1007
1008 r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault);
1009 if (r)
1010 return r;
1011
1012 /* Adjust VM size here.
1013 * Currently set to 4GB ((1 << 20) 4k pages).
1014 * Max GPUVM size for cayman and SI is 40 bits.
1015 */
1016 amdgpu_vm_adjust_size(adev, 64, 9, 1, 40);
1017
1018 /* Set the internal MC address mask
1019 * This is the max address of the GPU's
1020 * internal address space.
1021 */
1022 adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
1023
1024 r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(40));
1025 if (r) {
1026 pr_warn("No suitable DMA available\n");
1027 return r;
1028 }
1029 adev->need_swiotlb = drm_need_swiotlb(40);
1030
1031 r = gmc_v7_0_init_microcode(adev);
1032 if (r) {
1033 DRM_ERROR("Failed to load mc firmware!\n");
1034 return r;
1035 }
1036
1037 r = gmc_v7_0_mc_init(adev);
1038 if (r)
1039 return r;
1040
1041 amdgpu_gmc_get_vbios_allocations(adev);
1042
1043 /* Memory manager */
1044 r = amdgpu_bo_init(adev);
1045 if (r)
1046 return r;
1047
1048 r = gmc_v7_0_gart_init(adev);
1049 if (r)
1050 return r;
1051
1052 /*
1053 * number of VMs
1054 * VMID 0 is reserved for System
1055 * amdgpu graphics/compute will use VMIDs 1-7
1056 * amdkfd will use VMIDs 8-15
1057 */
1058 adev->vm_manager.first_kfd_vmid = 8;
1059 amdgpu_vm_manager_init(adev);
1060
1061 /* base offset of vram pages */
1062 if (adev->flags & AMD_IS_APU) {
1063 u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
1064
1065 tmp <<= 22;
1066 adev->vm_manager.vram_base_offset = tmp;
1067 } else {
1068 adev->vm_manager.vram_base_offset = 0;
1069 }
1070
1071 adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info),
1072 GFP_KERNEL);
1073 if (!adev->gmc.vm_fault_info)
1074 return -ENOMEM;
1075 atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1076
1077 return 0;
1078 }
1079
gmc_v7_0_sw_fini(void * handle)1080 static int gmc_v7_0_sw_fini(void *handle)
1081 {
1082 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1083
1084 amdgpu_gem_force_release(adev);
1085 amdgpu_vm_manager_fini(adev);
1086 kfree(adev->gmc.vm_fault_info);
1087 amdgpu_gart_table_vram_free(adev);
1088 amdgpu_bo_fini(adev);
1089 release_firmware(adev->gmc.fw);
1090 adev->gmc.fw = NULL;
1091
1092 return 0;
1093 }
1094
gmc_v7_0_hw_init(void * handle)1095 static int gmc_v7_0_hw_init(void *handle)
1096 {
1097 int r;
1098 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1099
1100 gmc_v7_0_init_golden_registers(adev);
1101
1102 gmc_v7_0_mc_program(adev);
1103
1104 if (!(adev->flags & AMD_IS_APU)) {
1105 r = gmc_v7_0_mc_load_microcode(adev);
1106 if (r) {
1107 DRM_ERROR("Failed to load MC firmware!\n");
1108 return r;
1109 }
1110 }
1111
1112 r = gmc_v7_0_gart_enable(adev);
1113 if (r)
1114 return r;
1115
1116 return r;
1117 }
1118
gmc_v7_0_hw_fini(void * handle)1119 static int gmc_v7_0_hw_fini(void *handle)
1120 {
1121 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1122
1123 amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1124 gmc_v7_0_gart_disable(adev);
1125
1126 return 0;
1127 }
1128
gmc_v7_0_suspend(void * handle)1129 static int gmc_v7_0_suspend(void *handle)
1130 {
1131 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1132
1133 gmc_v7_0_hw_fini(adev);
1134
1135 return 0;
1136 }
1137
gmc_v7_0_resume(void * handle)1138 static int gmc_v7_0_resume(void *handle)
1139 {
1140 int r;
1141 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1142
1143 r = gmc_v7_0_hw_init(adev);
1144 if (r)
1145 return r;
1146
1147 amdgpu_vmid_reset_all(adev);
1148
1149 return 0;
1150 }
1151
gmc_v7_0_is_idle(void * handle)1152 static bool gmc_v7_0_is_idle(void *handle)
1153 {
1154 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1155 u32 tmp = RREG32(mmSRBM_STATUS);
1156
1157 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1158 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1159 return false;
1160
1161 return true;
1162 }
1163
gmc_v7_0_wait_for_idle(void * handle)1164 static int gmc_v7_0_wait_for_idle(void *handle)
1165 {
1166 unsigned i;
1167 u32 tmp;
1168 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1169
1170 for (i = 0; i < adev->usec_timeout; i++) {
1171 /* read MC_STATUS */
1172 tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1173 SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1174 SRBM_STATUS__MCC_BUSY_MASK |
1175 SRBM_STATUS__MCD_BUSY_MASK |
1176 SRBM_STATUS__VMC_BUSY_MASK);
1177 if (!tmp)
1178 return 0;
1179 udelay(1);
1180 }
1181 return -ETIMEDOUT;
1182
1183 }
1184
gmc_v7_0_soft_reset(void * handle)1185 static int gmc_v7_0_soft_reset(void *handle)
1186 {
1187 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1188 u32 srbm_soft_reset = 0;
1189 u32 tmp = RREG32(mmSRBM_STATUS);
1190
1191 if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1192 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1193 SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1194
1195 if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1196 SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1197 if (!(adev->flags & AMD_IS_APU))
1198 srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1199 SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1200 }
1201
1202 if (srbm_soft_reset) {
1203 gmc_v7_0_mc_stop(adev);
1204 if (gmc_v7_0_wait_for_idle((void *)adev)) {
1205 dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1206 }
1207
1208
1209 tmp = RREG32(mmSRBM_SOFT_RESET);
1210 tmp |= srbm_soft_reset;
1211 dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1212 WREG32(mmSRBM_SOFT_RESET, tmp);
1213 tmp = RREG32(mmSRBM_SOFT_RESET);
1214
1215 udelay(50);
1216
1217 tmp &= ~srbm_soft_reset;
1218 WREG32(mmSRBM_SOFT_RESET, tmp);
1219 tmp = RREG32(mmSRBM_SOFT_RESET);
1220
1221 /* Wait a little for things to settle down */
1222 udelay(50);
1223
1224 gmc_v7_0_mc_resume(adev);
1225 udelay(50);
1226 }
1227
1228 return 0;
1229 }
1230
gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device * adev,struct amdgpu_irq_src * src,unsigned type,enum amdgpu_interrupt_state state)1231 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1232 struct amdgpu_irq_src *src,
1233 unsigned type,
1234 enum amdgpu_interrupt_state state)
1235 {
1236 u32 tmp;
1237 u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1238 VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1239 VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1240 VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1241 VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1242 VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1243
1244 switch (state) {
1245 case AMDGPU_IRQ_STATE_DISABLE:
1246 /* system context */
1247 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1248 tmp &= ~bits;
1249 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1250 /* VMs */
1251 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1252 tmp &= ~bits;
1253 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1254 break;
1255 case AMDGPU_IRQ_STATE_ENABLE:
1256 /* system context */
1257 tmp = RREG32(mmVM_CONTEXT0_CNTL);
1258 tmp |= bits;
1259 WREG32(mmVM_CONTEXT0_CNTL, tmp);
1260 /* VMs */
1261 tmp = RREG32(mmVM_CONTEXT1_CNTL);
1262 tmp |= bits;
1263 WREG32(mmVM_CONTEXT1_CNTL, tmp);
1264 break;
1265 default:
1266 break;
1267 }
1268
1269 return 0;
1270 }
1271
gmc_v7_0_process_interrupt(struct amdgpu_device * adev,struct amdgpu_irq_src * source,struct amdgpu_iv_entry * entry)1272 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev,
1273 struct amdgpu_irq_src *source,
1274 struct amdgpu_iv_entry *entry)
1275 {
1276 u32 addr, status, mc_client, vmid;
1277
1278 addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1279 status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1280 mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1281 /* reset addr and status */
1282 WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1283
1284 if (!addr && !status)
1285 return 0;
1286
1287 if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1288 gmc_v7_0_set_fault_enable_default(adev, false);
1289
1290 if (printk_ratelimit()) {
1291 dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1292 entry->src_id, entry->src_data[0]);
1293 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_ADDR 0x%08X\n",
1294 addr);
1295 dev_err(adev->dev, " VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1296 status);
1297 gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client,
1298 entry->pasid);
1299 }
1300
1301 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1302 VMID);
1303 if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid)
1304 && !atomic_read(&adev->gmc.vm_fault_info_updated)) {
1305 struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info;
1306 u32 protections = REG_GET_FIELD(status,
1307 VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1308 PROTECTIONS);
1309
1310 info->vmid = vmid;
1311 info->mc_id = REG_GET_FIELD(status,
1312 VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1313 MEMORY_CLIENT_ID);
1314 info->status = status;
1315 info->page_addr = addr;
1316 info->prot_valid = protections & 0x7 ? true : false;
1317 info->prot_read = protections & 0x8 ? true : false;
1318 info->prot_write = protections & 0x10 ? true : false;
1319 info->prot_exec = protections & 0x20 ? true : false;
1320 mb();
1321 atomic_set(&adev->gmc.vm_fault_info_updated, 1);
1322 }
1323
1324 return 0;
1325 }
1326
gmc_v7_0_set_clockgating_state(void * handle,enum amd_clockgating_state state)1327 static int gmc_v7_0_set_clockgating_state(void *handle,
1328 enum amd_clockgating_state state)
1329 {
1330 bool gate = false;
1331 struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1332
1333 if (state == AMD_CG_STATE_GATE)
1334 gate = true;
1335
1336 if (!(adev->flags & AMD_IS_APU)) {
1337 gmc_v7_0_enable_mc_mgcg(adev, gate);
1338 gmc_v7_0_enable_mc_ls(adev, gate);
1339 }
1340 gmc_v7_0_enable_bif_mgls(adev, gate);
1341 gmc_v7_0_enable_hdp_mgcg(adev, gate);
1342 gmc_v7_0_enable_hdp_ls(adev, gate);
1343
1344 return 0;
1345 }
1346
gmc_v7_0_set_powergating_state(void * handle,enum amd_powergating_state state)1347 static int gmc_v7_0_set_powergating_state(void *handle,
1348 enum amd_powergating_state state)
1349 {
1350 return 0;
1351 }
1352
1353 static const struct amd_ip_funcs gmc_v7_0_ip_funcs = {
1354 .name = "gmc_v7_0",
1355 .early_init = gmc_v7_0_early_init,
1356 .late_init = gmc_v7_0_late_init,
1357 .sw_init = gmc_v7_0_sw_init,
1358 .sw_fini = gmc_v7_0_sw_fini,
1359 .hw_init = gmc_v7_0_hw_init,
1360 .hw_fini = gmc_v7_0_hw_fini,
1361 .suspend = gmc_v7_0_suspend,
1362 .resume = gmc_v7_0_resume,
1363 .is_idle = gmc_v7_0_is_idle,
1364 .wait_for_idle = gmc_v7_0_wait_for_idle,
1365 .soft_reset = gmc_v7_0_soft_reset,
1366 .set_clockgating_state = gmc_v7_0_set_clockgating_state,
1367 .set_powergating_state = gmc_v7_0_set_powergating_state,
1368 };
1369
1370 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = {
1371 .flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb,
1372 .flush_gpu_tlb_pasid = gmc_v7_0_flush_gpu_tlb_pasid,
1373 .emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb,
1374 .emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping,
1375 .set_prt = gmc_v7_0_set_prt,
1376 .get_vm_pde = gmc_v7_0_get_vm_pde,
1377 .get_vm_pte = gmc_v7_0_get_vm_pte,
1378 .get_vbios_fb_size = gmc_v7_0_get_vbios_fb_size,
1379 };
1380
1381 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = {
1382 .set = gmc_v7_0_vm_fault_interrupt_state,
1383 .process = gmc_v7_0_process_interrupt,
1384 };
1385
gmc_v7_0_set_gmc_funcs(struct amdgpu_device * adev)1386 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev)
1387 {
1388 adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs;
1389 }
1390
gmc_v7_0_set_irq_funcs(struct amdgpu_device * adev)1391 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev)
1392 {
1393 adev->gmc.vm_fault.num_types = 1;
1394 adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs;
1395 }
1396
1397 const struct amdgpu_ip_block_version gmc_v7_0_ip_block =
1398 {
1399 .type = AMD_IP_BLOCK_TYPE_GMC,
1400 .major = 7,
1401 .minor = 0,
1402 .rev = 0,
1403 .funcs = &gmc_v7_0_ip_funcs,
1404 };
1405
1406 const struct amdgpu_ip_block_version gmc_v7_4_ip_block =
1407 {
1408 .type = AMD_IP_BLOCK_TYPE_GMC,
1409 .major = 7,
1410 .minor = 4,
1411 .rev = 0,
1412 .funcs = &gmc_v7_0_ip_funcs,
1413 };
1414