xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/gmc_v7_0.c (revision b830f94f)
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 
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 
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 
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  */
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  */
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 
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  */
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  */
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.aper_base = ((u64)RREG32(mmMC_VM_FB_OFFSET)) << 22;
386 		adev->gmc.aper_size = adev->gmc.real_vram_size;
387 	}
388 #endif
389 
390 	/* In case the PCI BAR is larger than the actual amount of vram */
391 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
392 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
393 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
394 
395 	/* set the gart size */
396 	if (amdgpu_gart_size == -1) {
397 		switch (adev->asic_type) {
398 		case CHIP_TOPAZ:     /* no MM engines */
399 		default:
400 			adev->gmc.gart_size = 256ULL << 20;
401 			break;
402 #ifdef CONFIG_DRM_AMDGPU_CIK
403 		case CHIP_BONAIRE: /* UVD, VCE do not support GPUVM */
404 		case CHIP_HAWAII:  /* UVD, VCE do not support GPUVM */
405 		case CHIP_KAVERI:  /* UVD, VCE do not support GPUVM */
406 		case CHIP_KABINI:  /* UVD, VCE do not support GPUVM */
407 		case CHIP_MULLINS: /* UVD, VCE do not support GPUVM */
408 			adev->gmc.gart_size = 1024ULL << 20;
409 			break;
410 #endif
411 		}
412 	} else {
413 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
414 	}
415 
416 	gmc_v7_0_vram_gtt_location(adev, &adev->gmc);
417 
418 	return 0;
419 }
420 
421 /*
422  * GART
423  * VMID 0 is the physical GPU addresses as used by the kernel.
424  * VMIDs 1-15 are used for userspace clients and are handled
425  * by the amdgpu vm/hsa code.
426  */
427 
428 /**
429  * gmc_v7_0_flush_gpu_tlb - gart tlb flush callback
430  *
431  * @adev: amdgpu_device pointer
432  * @vmid: vm instance to flush
433  *
434  * Flush the TLB for the requested page table (CIK).
435  */
436 static void gmc_v7_0_flush_gpu_tlb(struct amdgpu_device *adev,
437 				uint32_t vmid, uint32_t flush_type)
438 {
439 	/* bits 0-15 are the VM contexts0-15 */
440 	WREG32(mmVM_INVALIDATE_REQUEST, 1 << vmid);
441 }
442 
443 static uint64_t gmc_v7_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
444 					    unsigned vmid, uint64_t pd_addr)
445 {
446 	uint32_t reg;
447 
448 	if (vmid < 8)
449 		reg = mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + vmid;
450 	else
451 		reg = mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + vmid - 8;
452 	amdgpu_ring_emit_wreg(ring, reg, pd_addr >> 12);
453 
454 	/* bits 0-15 are the VM contexts0-15 */
455 	amdgpu_ring_emit_wreg(ring, mmVM_INVALIDATE_REQUEST, 1 << vmid);
456 
457 	return pd_addr;
458 }
459 
460 static void gmc_v7_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
461 					unsigned pasid)
462 {
463 	amdgpu_ring_emit_wreg(ring, mmIH_VMID_0_LUT + vmid, pasid);
464 }
465 
466 static uint64_t gmc_v7_0_get_vm_pte_flags(struct amdgpu_device *adev,
467 					  uint32_t flags)
468 {
469 	uint64_t pte_flag = 0;
470 
471 	if (flags & AMDGPU_VM_PAGE_READABLE)
472 		pte_flag |= AMDGPU_PTE_READABLE;
473 	if (flags & AMDGPU_VM_PAGE_WRITEABLE)
474 		pte_flag |= AMDGPU_PTE_WRITEABLE;
475 	if (flags & AMDGPU_VM_PAGE_PRT)
476 		pte_flag |= AMDGPU_PTE_PRT;
477 
478 	return pte_flag;
479 }
480 
481 static void gmc_v7_0_get_vm_pde(struct amdgpu_device *adev, int level,
482 				uint64_t *addr, uint64_t *flags)
483 {
484 	BUG_ON(*addr & 0xFFFFFF0000000FFFULL);
485 }
486 
487 /**
488  * gmc_v8_0_set_fault_enable_default - update VM fault handling
489  *
490  * @adev: amdgpu_device pointer
491  * @value: true redirects VM faults to the default page
492  */
493 static void gmc_v7_0_set_fault_enable_default(struct amdgpu_device *adev,
494 					      bool value)
495 {
496 	u32 tmp;
497 
498 	tmp = RREG32(mmVM_CONTEXT1_CNTL);
499 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
500 			    RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
501 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
502 			    DUMMY_PAGE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
503 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
504 			    PDE0_PROTECTION_FAULT_ENABLE_DEFAULT, value);
505 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
506 			    VALID_PROTECTION_FAULT_ENABLE_DEFAULT, value);
507 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
508 			    READ_PROTECTION_FAULT_ENABLE_DEFAULT, value);
509 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL,
510 			    WRITE_PROTECTION_FAULT_ENABLE_DEFAULT, value);
511 	WREG32(mmVM_CONTEXT1_CNTL, tmp);
512 }
513 
514 /**
515  * gmc_v7_0_set_prt - set PRT VM fault
516  *
517  * @adev: amdgpu_device pointer
518  * @enable: enable/disable VM fault handling for PRT
519  */
520 static void gmc_v7_0_set_prt(struct amdgpu_device *adev, bool enable)
521 {
522 	uint32_t tmp;
523 
524 	if (enable && !adev->gmc.prt_warning) {
525 		dev_warn(adev->dev, "Disabling VM faults because of PRT request!\n");
526 		adev->gmc.prt_warning = true;
527 	}
528 
529 	tmp = RREG32(mmVM_PRT_CNTL);
530 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
531 			    CB_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
532 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
533 			    CB_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
534 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
535 			    TC_DISABLE_READ_FAULT_ON_UNMAPPED_ACCESS, enable);
536 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
537 			    TC_DISABLE_WRITE_FAULT_ON_UNMAPPED_ACCESS, enable);
538 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
539 			    L2_CACHE_STORE_INVALID_ENTRIES, enable);
540 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
541 			    L1_TLB_STORE_INVALID_ENTRIES, enable);
542 	tmp = REG_SET_FIELD(tmp, VM_PRT_CNTL,
543 			    MASK_PDE0_FAULT, enable);
544 	WREG32(mmVM_PRT_CNTL, tmp);
545 
546 	if (enable) {
547 		uint32_t low = AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT;
548 		uint32_t high = adev->vm_manager.max_pfn -
549 			(AMDGPU_VA_RESERVED_SIZE >> AMDGPU_GPU_PAGE_SHIFT);
550 
551 		WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, low);
552 		WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, low);
553 		WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, low);
554 		WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, low);
555 		WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, high);
556 		WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, high);
557 		WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, high);
558 		WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, high);
559 	} else {
560 		WREG32(mmVM_PRT_APERTURE0_LOW_ADDR, 0xfffffff);
561 		WREG32(mmVM_PRT_APERTURE1_LOW_ADDR, 0xfffffff);
562 		WREG32(mmVM_PRT_APERTURE2_LOW_ADDR, 0xfffffff);
563 		WREG32(mmVM_PRT_APERTURE3_LOW_ADDR, 0xfffffff);
564 		WREG32(mmVM_PRT_APERTURE0_HIGH_ADDR, 0x0);
565 		WREG32(mmVM_PRT_APERTURE1_HIGH_ADDR, 0x0);
566 		WREG32(mmVM_PRT_APERTURE2_HIGH_ADDR, 0x0);
567 		WREG32(mmVM_PRT_APERTURE3_HIGH_ADDR, 0x0);
568 	}
569 }
570 
571 /**
572  * gmc_v7_0_gart_enable - gart enable
573  *
574  * @adev: amdgpu_device pointer
575  *
576  * This sets up the TLBs, programs the page tables for VMID0,
577  * sets up the hw for VMIDs 1-15 which are allocated on
578  * demand, and sets up the global locations for the LDS, GDS,
579  * and GPUVM for FSA64 clients (CIK).
580  * Returns 0 for success, errors for failure.
581  */
582 static int gmc_v7_0_gart_enable(struct amdgpu_device *adev)
583 {
584 	uint64_t table_addr;
585 	int r, i;
586 	u32 tmp, field;
587 
588 	if (adev->gart.bo == NULL) {
589 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
590 		return -EINVAL;
591 	}
592 	r = amdgpu_gart_table_vram_pin(adev);
593 	if (r)
594 		return r;
595 
596 	table_addr = amdgpu_bo_gpu_offset(adev->gart.bo);
597 
598 	/* Setup TLB control */
599 	tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
600 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 1);
601 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 1);
602 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_ACCESS_MODE, 3);
603 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 1);
604 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, SYSTEM_APERTURE_UNMAPPED_ACCESS, 0);
605 	WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
606 	/* Setup L2 cache */
607 	tmp = RREG32(mmVM_L2_CNTL);
608 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 1);
609 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_FRAGMENT_PROCESSING, 1);
610 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PTE_CACHE_LRU_UPDATE_BY_WRITE, 1);
611 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_PDE0_CACHE_LRU_UPDATE_BY_WRITE, 1);
612 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, EFFECTIVE_L2_QUEUE_SIZE, 7);
613 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, CONTEXT1_IDENTITY_ACCESS_MODE, 1);
614 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_DEFAULT_PAGE_OUT_TO_SYSTEM_MEMORY, 1);
615 	WREG32(mmVM_L2_CNTL, tmp);
616 	tmp = REG_SET_FIELD(0, VM_L2_CNTL2, INVALIDATE_ALL_L1_TLBS, 1);
617 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL2, INVALIDATE_L2_CACHE, 1);
618 	WREG32(mmVM_L2_CNTL2, tmp);
619 
620 	field = adev->vm_manager.fragment_size;
621 	tmp = RREG32(mmVM_L2_CNTL3);
622 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_ASSOCIATIVITY, 1);
623 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, BANK_SELECT, field);
624 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL3, L2_CACHE_BIGK_FRAGMENT_SIZE, field);
625 	WREG32(mmVM_L2_CNTL3, tmp);
626 	/* setup context0 */
627 	WREG32(mmVM_CONTEXT0_PAGE_TABLE_START_ADDR, adev->gmc.gart_start >> 12);
628 	WREG32(mmVM_CONTEXT0_PAGE_TABLE_END_ADDR, adev->gmc.gart_end >> 12);
629 	WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR, table_addr >> 12);
630 	WREG32(mmVM_CONTEXT0_PROTECTION_FAULT_DEFAULT_ADDR,
631 			(u32)(adev->dummy_page_addr >> 12));
632 	WREG32(mmVM_CONTEXT0_CNTL2, 0);
633 	tmp = RREG32(mmVM_CONTEXT0_CNTL);
634 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, ENABLE_CONTEXT, 1);
635 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, PAGE_TABLE_DEPTH, 0);
636 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT0_CNTL, RANGE_PROTECTION_FAULT_ENABLE_DEFAULT, 1);
637 	WREG32(mmVM_CONTEXT0_CNTL, tmp);
638 
639 	WREG32(0x575, 0);
640 	WREG32(0x576, 0);
641 	WREG32(0x577, 0);
642 
643 	/* empty context1-15 */
644 	/* FIXME start with 4G, once using 2 level pt switch to full
645 	 * vm size space
646 	 */
647 	/* set vm size, must be a multiple of 4 */
648 	WREG32(mmVM_CONTEXT1_PAGE_TABLE_START_ADDR, 0);
649 	WREG32(mmVM_CONTEXT1_PAGE_TABLE_END_ADDR, adev->vm_manager.max_pfn - 1);
650 	for (i = 1; i < 16; i++) {
651 		if (i < 8)
652 			WREG32(mmVM_CONTEXT0_PAGE_TABLE_BASE_ADDR + i,
653 			       table_addr >> 12);
654 		else
655 			WREG32(mmVM_CONTEXT8_PAGE_TABLE_BASE_ADDR + i - 8,
656 			       table_addr >> 12);
657 	}
658 
659 	/* enable context1-15 */
660 	WREG32(mmVM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR,
661 	       (u32)(adev->dummy_page_addr >> 12));
662 	WREG32(mmVM_CONTEXT1_CNTL2, 4);
663 	tmp = RREG32(mmVM_CONTEXT1_CNTL);
664 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, ENABLE_CONTEXT, 1);
665 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_DEPTH, 1);
666 	tmp = REG_SET_FIELD(tmp, VM_CONTEXT1_CNTL, PAGE_TABLE_BLOCK_SIZE,
667 			    adev->vm_manager.block_size - 9);
668 	WREG32(mmVM_CONTEXT1_CNTL, tmp);
669 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS)
670 		gmc_v7_0_set_fault_enable_default(adev, false);
671 	else
672 		gmc_v7_0_set_fault_enable_default(adev, true);
673 
674 	if (adev->asic_type == CHIP_KAVERI) {
675 		tmp = RREG32(mmCHUB_CONTROL);
676 		tmp &= ~BYPASS_VM;
677 		WREG32(mmCHUB_CONTROL, tmp);
678 	}
679 
680 	gmc_v7_0_flush_gpu_tlb(adev, 0, 0);
681 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
682 		 (unsigned)(adev->gmc.gart_size >> 20),
683 		 (unsigned long long)table_addr);
684 	adev->gart.ready = true;
685 	return 0;
686 }
687 
688 static int gmc_v7_0_gart_init(struct amdgpu_device *adev)
689 {
690 	int r;
691 
692 	if (adev->gart.bo) {
693 		WARN(1, "R600 PCIE GART already initialized\n");
694 		return 0;
695 	}
696 	/* Initialize common gart structure */
697 	r = amdgpu_gart_init(adev);
698 	if (r)
699 		return r;
700 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
701 	adev->gart.gart_pte_flags = 0;
702 	return amdgpu_gart_table_vram_alloc(adev);
703 }
704 
705 /**
706  * gmc_v7_0_gart_disable - gart disable
707  *
708  * @adev: amdgpu_device pointer
709  *
710  * This disables all VM page table (CIK).
711  */
712 static void gmc_v7_0_gart_disable(struct amdgpu_device *adev)
713 {
714 	u32 tmp;
715 
716 	/* Disable all tables */
717 	WREG32(mmVM_CONTEXT0_CNTL, 0);
718 	WREG32(mmVM_CONTEXT1_CNTL, 0);
719 	/* Setup TLB control */
720 	tmp = RREG32(mmMC_VM_MX_L1_TLB_CNTL);
721 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_TLB, 0);
722 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_L1_FRAGMENT_PROCESSING, 0);
723 	tmp = REG_SET_FIELD(tmp, MC_VM_MX_L1_TLB_CNTL, ENABLE_ADVANCED_DRIVER_MODEL, 0);
724 	WREG32(mmMC_VM_MX_L1_TLB_CNTL, tmp);
725 	/* Setup L2 cache */
726 	tmp = RREG32(mmVM_L2_CNTL);
727 	tmp = REG_SET_FIELD(tmp, VM_L2_CNTL, ENABLE_L2_CACHE, 0);
728 	WREG32(mmVM_L2_CNTL, tmp);
729 	WREG32(mmVM_L2_CNTL2, 0);
730 	amdgpu_gart_table_vram_unpin(adev);
731 }
732 
733 /**
734  * gmc_v7_0_vm_decode_fault - print human readable fault info
735  *
736  * @adev: amdgpu_device pointer
737  * @status: VM_CONTEXT1_PROTECTION_FAULT_STATUS register value
738  * @addr: VM_CONTEXT1_PROTECTION_FAULT_ADDR register value
739  *
740  * Print human readable fault information (CIK).
741  */
742 static void gmc_v7_0_vm_decode_fault(struct amdgpu_device *adev, u32 status,
743 				     u32 addr, u32 mc_client, unsigned pasid)
744 {
745 	u32 vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS, VMID);
746 	u32 protections = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
747 					PROTECTIONS);
748 	char block[5] = { mc_client >> 24, (mc_client >> 16) & 0xff,
749 		(mc_client >> 8) & 0xff, mc_client & 0xff, 0 };
750 	u32 mc_id;
751 
752 	mc_id = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
753 			      MEMORY_CLIENT_ID);
754 
755 	dev_err(adev->dev, "VM fault (0x%02x, vmid %d, pasid %d) at page %u, %s from '%s' (0x%08x) (%d)\n",
756 	       protections, vmid, pasid, addr,
757 	       REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
758 			     MEMORY_CLIENT_RW) ?
759 	       "write" : "read", block, mc_client, mc_id);
760 }
761 
762 
763 static const u32 mc_cg_registers[] = {
764 	mmMC_HUB_MISC_HUB_CG,
765 	mmMC_HUB_MISC_SIP_CG,
766 	mmMC_HUB_MISC_VM_CG,
767 	mmMC_XPB_CLK_GAT,
768 	mmATC_MISC_CG,
769 	mmMC_CITF_MISC_WR_CG,
770 	mmMC_CITF_MISC_RD_CG,
771 	mmMC_CITF_MISC_VM_CG,
772 	mmVM_L2_CG,
773 };
774 
775 static const u32 mc_cg_ls_en[] = {
776 	MC_HUB_MISC_HUB_CG__MEM_LS_ENABLE_MASK,
777 	MC_HUB_MISC_SIP_CG__MEM_LS_ENABLE_MASK,
778 	MC_HUB_MISC_VM_CG__MEM_LS_ENABLE_MASK,
779 	MC_XPB_CLK_GAT__MEM_LS_ENABLE_MASK,
780 	ATC_MISC_CG__MEM_LS_ENABLE_MASK,
781 	MC_CITF_MISC_WR_CG__MEM_LS_ENABLE_MASK,
782 	MC_CITF_MISC_RD_CG__MEM_LS_ENABLE_MASK,
783 	MC_CITF_MISC_VM_CG__MEM_LS_ENABLE_MASK,
784 	VM_L2_CG__MEM_LS_ENABLE_MASK,
785 };
786 
787 static const u32 mc_cg_en[] = {
788 	MC_HUB_MISC_HUB_CG__ENABLE_MASK,
789 	MC_HUB_MISC_SIP_CG__ENABLE_MASK,
790 	MC_HUB_MISC_VM_CG__ENABLE_MASK,
791 	MC_XPB_CLK_GAT__ENABLE_MASK,
792 	ATC_MISC_CG__ENABLE_MASK,
793 	MC_CITF_MISC_WR_CG__ENABLE_MASK,
794 	MC_CITF_MISC_RD_CG__ENABLE_MASK,
795 	MC_CITF_MISC_VM_CG__ENABLE_MASK,
796 	VM_L2_CG__ENABLE_MASK,
797 };
798 
799 static void gmc_v7_0_enable_mc_ls(struct amdgpu_device *adev,
800 				  bool enable)
801 {
802 	int i;
803 	u32 orig, data;
804 
805 	for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
806 		orig = data = RREG32(mc_cg_registers[i]);
807 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_LS))
808 			data |= mc_cg_ls_en[i];
809 		else
810 			data &= ~mc_cg_ls_en[i];
811 		if (data != orig)
812 			WREG32(mc_cg_registers[i], data);
813 	}
814 }
815 
816 static void gmc_v7_0_enable_mc_mgcg(struct amdgpu_device *adev,
817 				    bool enable)
818 {
819 	int i;
820 	u32 orig, data;
821 
822 	for (i = 0; i < ARRAY_SIZE(mc_cg_registers); i++) {
823 		orig = data = RREG32(mc_cg_registers[i]);
824 		if (enable && (adev->cg_flags & AMD_CG_SUPPORT_MC_MGCG))
825 			data |= mc_cg_en[i];
826 		else
827 			data &= ~mc_cg_en[i];
828 		if (data != orig)
829 			WREG32(mc_cg_registers[i], data);
830 	}
831 }
832 
833 static void gmc_v7_0_enable_bif_mgls(struct amdgpu_device *adev,
834 				     bool enable)
835 {
836 	u32 orig, data;
837 
838 	orig = data = RREG32_PCIE(ixPCIE_CNTL2);
839 
840 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_BIF_LS)) {
841 		data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 1);
842 		data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 1);
843 		data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 1);
844 		data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 1);
845 	} else {
846 		data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_LS_EN, 0);
847 		data = REG_SET_FIELD(data, PCIE_CNTL2, MST_MEM_LS_EN, 0);
848 		data = REG_SET_FIELD(data, PCIE_CNTL2, REPLAY_MEM_LS_EN, 0);
849 		data = REG_SET_FIELD(data, PCIE_CNTL2, SLV_MEM_AGGRESSIVE_LS_EN, 0);
850 	}
851 
852 	if (orig != data)
853 		WREG32_PCIE(ixPCIE_CNTL2, data);
854 }
855 
856 static void gmc_v7_0_enable_hdp_mgcg(struct amdgpu_device *adev,
857 				     bool enable)
858 {
859 	u32 orig, data;
860 
861 	orig = data = RREG32(mmHDP_HOST_PATH_CNTL);
862 
863 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_MGCG))
864 		data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 0);
865 	else
866 		data = REG_SET_FIELD(data, HDP_HOST_PATH_CNTL, CLOCK_GATING_DIS, 1);
867 
868 	if (orig != data)
869 		WREG32(mmHDP_HOST_PATH_CNTL, data);
870 }
871 
872 static void gmc_v7_0_enable_hdp_ls(struct amdgpu_device *adev,
873 				   bool enable)
874 {
875 	u32 orig, data;
876 
877 	orig = data = RREG32(mmHDP_MEM_POWER_LS);
878 
879 	if (enable && (adev->cg_flags & AMD_CG_SUPPORT_HDP_LS))
880 		data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 1);
881 	else
882 		data = REG_SET_FIELD(data, HDP_MEM_POWER_LS, LS_ENABLE, 0);
883 
884 	if (orig != data)
885 		WREG32(mmHDP_MEM_POWER_LS, data);
886 }
887 
888 static int gmc_v7_0_convert_vram_type(int mc_seq_vram_type)
889 {
890 	switch (mc_seq_vram_type) {
891 	case MC_SEQ_MISC0__MT__GDDR1:
892 		return AMDGPU_VRAM_TYPE_GDDR1;
893 	case MC_SEQ_MISC0__MT__DDR2:
894 		return AMDGPU_VRAM_TYPE_DDR2;
895 	case MC_SEQ_MISC0__MT__GDDR3:
896 		return AMDGPU_VRAM_TYPE_GDDR3;
897 	case MC_SEQ_MISC0__MT__GDDR4:
898 		return AMDGPU_VRAM_TYPE_GDDR4;
899 	case MC_SEQ_MISC0__MT__GDDR5:
900 		return AMDGPU_VRAM_TYPE_GDDR5;
901 	case MC_SEQ_MISC0__MT__HBM:
902 		return AMDGPU_VRAM_TYPE_HBM;
903 	case MC_SEQ_MISC0__MT__DDR3:
904 		return AMDGPU_VRAM_TYPE_DDR3;
905 	default:
906 		return AMDGPU_VRAM_TYPE_UNKNOWN;
907 	}
908 }
909 
910 static int gmc_v7_0_early_init(void *handle)
911 {
912 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
913 
914 	gmc_v7_0_set_gmc_funcs(adev);
915 	gmc_v7_0_set_irq_funcs(adev);
916 
917 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
918 	adev->gmc.shared_aperture_end =
919 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
920 	adev->gmc.private_aperture_start =
921 		adev->gmc.shared_aperture_end + 1;
922 	adev->gmc.private_aperture_end =
923 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
924 
925 	return 0;
926 }
927 
928 static int gmc_v7_0_late_init(void *handle)
929 {
930 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
931 
932 	amdgpu_bo_late_init(adev);
933 
934 	if (amdgpu_vm_fault_stop != AMDGPU_VM_FAULT_STOP_ALWAYS)
935 		return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
936 	else
937 		return 0;
938 }
939 
940 static unsigned gmc_v7_0_get_vbios_fb_size(struct amdgpu_device *adev)
941 {
942 	u32 d1vga_control = RREG32(mmD1VGA_CONTROL);
943 	unsigned size;
944 
945 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
946 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
947 	} else {
948 		u32 viewport = RREG32(mmVIEWPORT_SIZE);
949 		size = (REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
950 			REG_GET_FIELD(viewport, VIEWPORT_SIZE, VIEWPORT_WIDTH) *
951 			4);
952 	}
953 	/* return 0 if the pre-OS buffer uses up most of vram */
954 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
955 		return 0;
956 	return size;
957 }
958 
959 static int gmc_v7_0_sw_init(void *handle)
960 {
961 	int r;
962 	int dma_bits;
963 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
964 
965 	if (adev->flags & AMD_IS_APU) {
966 		adev->gmc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
967 	} else {
968 		u32 tmp = RREG32(mmMC_SEQ_MISC0);
969 		tmp &= MC_SEQ_MISC0__MT__MASK;
970 		adev->gmc.vram_type = gmc_v7_0_convert_vram_type(tmp);
971 	}
972 
973 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_PAGE_INV_FAULT, &adev->gmc.vm_fault);
974 	if (r)
975 		return r;
976 
977 	r = amdgpu_irq_add_id(adev, AMDGPU_IRQ_CLIENTID_LEGACY, VISLANDS30_IV_SRCID_GFX_MEM_PROT_FAULT, &adev->gmc.vm_fault);
978 	if (r)
979 		return r;
980 
981 	/* Adjust VM size here.
982 	 * Currently set to 4GB ((1 << 20) 4k pages).
983 	 * Max GPUVM size for cayman and SI is 40 bits.
984 	 */
985 	amdgpu_vm_adjust_size(adev, 64, 9, 1, 40);
986 
987 	/* Set the internal MC address mask
988 	 * This is the max address of the GPU's
989 	 * internal address space.
990 	 */
991 	adev->gmc.mc_mask = 0xffffffffffULL; /* 40 bit MC */
992 
993 	/* set DMA mask + need_dma32 flags.
994 	 * PCIE - can handle 40-bits.
995 	 * IGP - can handle 40-bits
996 	 * PCI - dma32 for legacy pci gart, 40 bits on newer asics
997 	 */
998 	adev->need_dma32 = false;
999 	dma_bits = adev->need_dma32 ? 32 : 40;
1000 	r = pci_set_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
1001 	if (r) {
1002 		adev->need_dma32 = true;
1003 		dma_bits = 32;
1004 		pr_warn("amdgpu: No suitable DMA available\n");
1005 	}
1006 	r = pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(dma_bits));
1007 	if (r) {
1008 		pci_set_consistent_dma_mask(adev->pdev, DMA_BIT_MASK(32));
1009 		pr_warn("amdgpu: No coherent DMA available\n");
1010 	}
1011 	adev->need_swiotlb = drm_need_swiotlb(dma_bits);
1012 
1013 	r = gmc_v7_0_init_microcode(adev);
1014 	if (r) {
1015 		DRM_ERROR("Failed to load mc firmware!\n");
1016 		return r;
1017 	}
1018 
1019 	r = gmc_v7_0_mc_init(adev);
1020 	if (r)
1021 		return r;
1022 
1023 	adev->gmc.stolen_size = gmc_v7_0_get_vbios_fb_size(adev);
1024 
1025 	/* Memory manager */
1026 	r = amdgpu_bo_init(adev);
1027 	if (r)
1028 		return r;
1029 
1030 	r = gmc_v7_0_gart_init(adev);
1031 	if (r)
1032 		return r;
1033 
1034 	/*
1035 	 * number of VMs
1036 	 * VMID 0 is reserved for System
1037 	 * amdgpu graphics/compute will use VMIDs 1-7
1038 	 * amdkfd will use VMIDs 8-15
1039 	 */
1040 	adev->vm_manager.id_mgr[0].num_ids = AMDGPU_NUM_OF_VMIDS;
1041 	amdgpu_vm_manager_init(adev);
1042 
1043 	/* base offset of vram pages */
1044 	if (adev->flags & AMD_IS_APU) {
1045 		u64 tmp = RREG32(mmMC_VM_FB_OFFSET);
1046 
1047 		tmp <<= 22;
1048 		adev->vm_manager.vram_base_offset = tmp;
1049 	} else {
1050 		adev->vm_manager.vram_base_offset = 0;
1051 	}
1052 
1053 	adev->gmc.vm_fault_info = kmalloc(sizeof(struct kfd_vm_fault_info),
1054 					GFP_KERNEL);
1055 	if (!adev->gmc.vm_fault_info)
1056 		return -ENOMEM;
1057 	atomic_set(&adev->gmc.vm_fault_info_updated, 0);
1058 
1059 	return 0;
1060 }
1061 
1062 static int gmc_v7_0_sw_fini(void *handle)
1063 {
1064 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1065 
1066 	amdgpu_gem_force_release(adev);
1067 	amdgpu_vm_manager_fini(adev);
1068 	kfree(adev->gmc.vm_fault_info);
1069 	amdgpu_gart_table_vram_free(adev);
1070 	amdgpu_bo_fini(adev);
1071 	amdgpu_gart_fini(adev);
1072 	release_firmware(adev->gmc.fw);
1073 	adev->gmc.fw = NULL;
1074 
1075 	return 0;
1076 }
1077 
1078 static int gmc_v7_0_hw_init(void *handle)
1079 {
1080 	int r;
1081 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1082 
1083 	gmc_v7_0_init_golden_registers(adev);
1084 
1085 	gmc_v7_0_mc_program(adev);
1086 
1087 	if (!(adev->flags & AMD_IS_APU)) {
1088 		r = gmc_v7_0_mc_load_microcode(adev);
1089 		if (r) {
1090 			DRM_ERROR("Failed to load MC firmware!\n");
1091 			return r;
1092 		}
1093 	}
1094 
1095 	r = gmc_v7_0_gart_enable(adev);
1096 	if (r)
1097 		return r;
1098 
1099 	return r;
1100 }
1101 
1102 static int gmc_v7_0_hw_fini(void *handle)
1103 {
1104 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1105 
1106 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1107 	gmc_v7_0_gart_disable(adev);
1108 
1109 	return 0;
1110 }
1111 
1112 static int gmc_v7_0_suspend(void *handle)
1113 {
1114 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1115 
1116 	gmc_v7_0_hw_fini(adev);
1117 
1118 	return 0;
1119 }
1120 
1121 static int gmc_v7_0_resume(void *handle)
1122 {
1123 	int r;
1124 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1125 
1126 	r = gmc_v7_0_hw_init(adev);
1127 	if (r)
1128 		return r;
1129 
1130 	amdgpu_vmid_reset_all(adev);
1131 
1132 	return 0;
1133 }
1134 
1135 static bool gmc_v7_0_is_idle(void *handle)
1136 {
1137 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1138 	u32 tmp = RREG32(mmSRBM_STATUS);
1139 
1140 	if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1141 		   SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK | SRBM_STATUS__VMC_BUSY_MASK))
1142 		return false;
1143 
1144 	return true;
1145 }
1146 
1147 static int gmc_v7_0_wait_for_idle(void *handle)
1148 {
1149 	unsigned i;
1150 	u32 tmp;
1151 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1152 
1153 	for (i = 0; i < adev->usec_timeout; i++) {
1154 		/* read MC_STATUS */
1155 		tmp = RREG32(mmSRBM_STATUS) & (SRBM_STATUS__MCB_BUSY_MASK |
1156 					       SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1157 					       SRBM_STATUS__MCC_BUSY_MASK |
1158 					       SRBM_STATUS__MCD_BUSY_MASK |
1159 					       SRBM_STATUS__VMC_BUSY_MASK);
1160 		if (!tmp)
1161 			return 0;
1162 		udelay(1);
1163 	}
1164 	return -ETIMEDOUT;
1165 
1166 }
1167 
1168 static int gmc_v7_0_soft_reset(void *handle)
1169 {
1170 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1171 	u32 srbm_soft_reset = 0;
1172 	u32 tmp = RREG32(mmSRBM_STATUS);
1173 
1174 	if (tmp & SRBM_STATUS__VMC_BUSY_MASK)
1175 		srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1176 						SRBM_SOFT_RESET, SOFT_RESET_VMC, 1);
1177 
1178 	if (tmp & (SRBM_STATUS__MCB_BUSY_MASK | SRBM_STATUS__MCB_NON_DISPLAY_BUSY_MASK |
1179 		   SRBM_STATUS__MCC_BUSY_MASK | SRBM_STATUS__MCD_BUSY_MASK)) {
1180 		if (!(adev->flags & AMD_IS_APU))
1181 			srbm_soft_reset = REG_SET_FIELD(srbm_soft_reset,
1182 							SRBM_SOFT_RESET, SOFT_RESET_MC, 1);
1183 	}
1184 
1185 	if (srbm_soft_reset) {
1186 		gmc_v7_0_mc_stop(adev);
1187 		if (gmc_v7_0_wait_for_idle((void *)adev)) {
1188 			dev_warn(adev->dev, "Wait for GMC idle timed out !\n");
1189 		}
1190 
1191 
1192 		tmp = RREG32(mmSRBM_SOFT_RESET);
1193 		tmp |= srbm_soft_reset;
1194 		dev_info(adev->dev, "SRBM_SOFT_RESET=0x%08X\n", tmp);
1195 		WREG32(mmSRBM_SOFT_RESET, tmp);
1196 		tmp = RREG32(mmSRBM_SOFT_RESET);
1197 
1198 		udelay(50);
1199 
1200 		tmp &= ~srbm_soft_reset;
1201 		WREG32(mmSRBM_SOFT_RESET, tmp);
1202 		tmp = RREG32(mmSRBM_SOFT_RESET);
1203 
1204 		/* Wait a little for things to settle down */
1205 		udelay(50);
1206 
1207 		gmc_v7_0_mc_resume(adev);
1208 		udelay(50);
1209 	}
1210 
1211 	return 0;
1212 }
1213 
1214 static int gmc_v7_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
1215 					     struct amdgpu_irq_src *src,
1216 					     unsigned type,
1217 					     enum amdgpu_interrupt_state state)
1218 {
1219 	u32 tmp;
1220 	u32 bits = (VM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1221 		    VM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1222 		    VM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1223 		    VM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1224 		    VM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
1225 		    VM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK);
1226 
1227 	switch (state) {
1228 	case AMDGPU_IRQ_STATE_DISABLE:
1229 		/* system context */
1230 		tmp = RREG32(mmVM_CONTEXT0_CNTL);
1231 		tmp &= ~bits;
1232 		WREG32(mmVM_CONTEXT0_CNTL, tmp);
1233 		/* VMs */
1234 		tmp = RREG32(mmVM_CONTEXT1_CNTL);
1235 		tmp &= ~bits;
1236 		WREG32(mmVM_CONTEXT1_CNTL, tmp);
1237 		break;
1238 	case AMDGPU_IRQ_STATE_ENABLE:
1239 		/* system context */
1240 		tmp = RREG32(mmVM_CONTEXT0_CNTL);
1241 		tmp |= bits;
1242 		WREG32(mmVM_CONTEXT0_CNTL, tmp);
1243 		/* VMs */
1244 		tmp = RREG32(mmVM_CONTEXT1_CNTL);
1245 		tmp |= bits;
1246 		WREG32(mmVM_CONTEXT1_CNTL, tmp);
1247 		break;
1248 	default:
1249 		break;
1250 	}
1251 
1252 	return 0;
1253 }
1254 
1255 static int gmc_v7_0_process_interrupt(struct amdgpu_device *adev,
1256 				      struct amdgpu_irq_src *source,
1257 				      struct amdgpu_iv_entry *entry)
1258 {
1259 	u32 addr, status, mc_client, vmid;
1260 
1261 	addr = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_ADDR);
1262 	status = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_STATUS);
1263 	mc_client = RREG32(mmVM_CONTEXT1_PROTECTION_FAULT_MCCLIENT);
1264 	/* reset addr and status */
1265 	WREG32_P(mmVM_CONTEXT1_CNTL2, 1, ~1);
1266 
1267 	if (!addr && !status)
1268 		return 0;
1269 
1270 	if (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_FIRST)
1271 		gmc_v7_0_set_fault_enable_default(adev, false);
1272 
1273 	if (printk_ratelimit()) {
1274 		dev_err(adev->dev, "GPU fault detected: %d 0x%08x\n",
1275 			entry->src_id, entry->src_data[0]);
1276 		dev_err(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_ADDR   0x%08X\n",
1277 			addr);
1278 		dev_err(adev->dev, "  VM_CONTEXT1_PROTECTION_FAULT_STATUS 0x%08X\n",
1279 			status);
1280 		gmc_v7_0_vm_decode_fault(adev, status, addr, mc_client,
1281 					 entry->pasid);
1282 	}
1283 
1284 	vmid = REG_GET_FIELD(status, VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1285 			     VMID);
1286 	if (amdgpu_amdkfd_is_kfd_vmid(adev, vmid)
1287 		&& !atomic_read(&adev->gmc.vm_fault_info_updated)) {
1288 		struct kfd_vm_fault_info *info = adev->gmc.vm_fault_info;
1289 		u32 protections = REG_GET_FIELD(status,
1290 					VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1291 					PROTECTIONS);
1292 
1293 		info->vmid = vmid;
1294 		info->mc_id = REG_GET_FIELD(status,
1295 					    VM_CONTEXT1_PROTECTION_FAULT_STATUS,
1296 					    MEMORY_CLIENT_ID);
1297 		info->status = status;
1298 		info->page_addr = addr;
1299 		info->prot_valid = protections & 0x7 ? true : false;
1300 		info->prot_read = protections & 0x8 ? true : false;
1301 		info->prot_write = protections & 0x10 ? true : false;
1302 		info->prot_exec = protections & 0x20 ? true : false;
1303 		mb();
1304 		atomic_set(&adev->gmc.vm_fault_info_updated, 1);
1305 	}
1306 
1307 	return 0;
1308 }
1309 
1310 static int gmc_v7_0_set_clockgating_state(void *handle,
1311 					  enum amd_clockgating_state state)
1312 {
1313 	bool gate = false;
1314 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1315 
1316 	if (state == AMD_CG_STATE_GATE)
1317 		gate = true;
1318 
1319 	if (!(adev->flags & AMD_IS_APU)) {
1320 		gmc_v7_0_enable_mc_mgcg(adev, gate);
1321 		gmc_v7_0_enable_mc_ls(adev, gate);
1322 	}
1323 	gmc_v7_0_enable_bif_mgls(adev, gate);
1324 	gmc_v7_0_enable_hdp_mgcg(adev, gate);
1325 	gmc_v7_0_enable_hdp_ls(adev, gate);
1326 
1327 	return 0;
1328 }
1329 
1330 static int gmc_v7_0_set_powergating_state(void *handle,
1331 					  enum amd_powergating_state state)
1332 {
1333 	return 0;
1334 }
1335 
1336 static const struct amd_ip_funcs gmc_v7_0_ip_funcs = {
1337 	.name = "gmc_v7_0",
1338 	.early_init = gmc_v7_0_early_init,
1339 	.late_init = gmc_v7_0_late_init,
1340 	.sw_init = gmc_v7_0_sw_init,
1341 	.sw_fini = gmc_v7_0_sw_fini,
1342 	.hw_init = gmc_v7_0_hw_init,
1343 	.hw_fini = gmc_v7_0_hw_fini,
1344 	.suspend = gmc_v7_0_suspend,
1345 	.resume = gmc_v7_0_resume,
1346 	.is_idle = gmc_v7_0_is_idle,
1347 	.wait_for_idle = gmc_v7_0_wait_for_idle,
1348 	.soft_reset = gmc_v7_0_soft_reset,
1349 	.set_clockgating_state = gmc_v7_0_set_clockgating_state,
1350 	.set_powergating_state = gmc_v7_0_set_powergating_state,
1351 };
1352 
1353 static const struct amdgpu_gmc_funcs gmc_v7_0_gmc_funcs = {
1354 	.flush_gpu_tlb = gmc_v7_0_flush_gpu_tlb,
1355 	.emit_flush_gpu_tlb = gmc_v7_0_emit_flush_gpu_tlb,
1356 	.emit_pasid_mapping = gmc_v7_0_emit_pasid_mapping,
1357 	.set_prt = gmc_v7_0_set_prt,
1358 	.get_vm_pte_flags = gmc_v7_0_get_vm_pte_flags,
1359 	.get_vm_pde = gmc_v7_0_get_vm_pde
1360 };
1361 
1362 static const struct amdgpu_irq_src_funcs gmc_v7_0_irq_funcs = {
1363 	.set = gmc_v7_0_vm_fault_interrupt_state,
1364 	.process = gmc_v7_0_process_interrupt,
1365 };
1366 
1367 static void gmc_v7_0_set_gmc_funcs(struct amdgpu_device *adev)
1368 {
1369 	adev->gmc.gmc_funcs = &gmc_v7_0_gmc_funcs;
1370 }
1371 
1372 static void gmc_v7_0_set_irq_funcs(struct amdgpu_device *adev)
1373 {
1374 	adev->gmc.vm_fault.num_types = 1;
1375 	adev->gmc.vm_fault.funcs = &gmc_v7_0_irq_funcs;
1376 }
1377 
1378 const struct amdgpu_ip_block_version gmc_v7_0_ip_block =
1379 {
1380 	.type = AMD_IP_BLOCK_TYPE_GMC,
1381 	.major = 7,
1382 	.minor = 0,
1383 	.rev = 0,
1384 	.funcs = &gmc_v7_0_ip_funcs,
1385 };
1386 
1387 const struct amdgpu_ip_block_version gmc_v7_4_ip_block =
1388 {
1389 	.type = AMD_IP_BLOCK_TYPE_GMC,
1390 	.major = 7,
1391 	.minor = 4,
1392 	.rev = 0,
1393 	.funcs = &gmc_v7_0_ip_funcs,
1394 };
1395