1 /*
2  * Copyright 2019 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 <linux/pci.h>
25 #include "amdgpu.h"
26 #include "amdgpu_atomfirmware.h"
27 #include "gmc_v10_0.h"
28 
29 #include "hdp/hdp_5_0_0_offset.h"
30 #include "hdp/hdp_5_0_0_sh_mask.h"
31 #include "gc/gc_10_1_0_sh_mask.h"
32 #include "mmhub/mmhub_2_0_0_sh_mask.h"
33 #include "dcn/dcn_2_0_0_offset.h"
34 #include "dcn/dcn_2_0_0_sh_mask.h"
35 #include "oss/osssys_5_0_0_offset.h"
36 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
37 #include "navi10_enum.h"
38 
39 #include "soc15.h"
40 #include "soc15_common.h"
41 
42 #include "nbio_v2_3.h"
43 
44 #include "gfxhub_v2_0.h"
45 #include "mmhub_v2_0.h"
46 #include "athub_v2_0.h"
47 /* XXX Move this macro to navi10 header file, which is like vid.h for VI.*/
48 #define AMDGPU_NUM_OF_VMIDS			8
49 
50 #if 0
51 static const struct soc15_reg_golden golden_settings_navi10_hdp[] =
52 {
53 	/* TODO add golden setting for hdp */
54 };
55 #endif
56 
57 static int
58 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
59 				   struct amdgpu_irq_src *src, unsigned type,
60 				   enum amdgpu_interrupt_state state)
61 {
62 	struct amdgpu_vmhub *hub;
63 	u32 tmp, reg, bits[AMDGPU_MAX_VMHUBS], i;
64 
65 	bits[AMDGPU_GFXHUB_0] = GCVM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
66 		GCVM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
67 		GCVM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
68 		GCVM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
69 		GCVM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
70 		GCVM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
71 		GCVM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
72 
73 	bits[AMDGPU_MMHUB_0] = MMVM_CONTEXT1_CNTL__RANGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
74 		MMVM_CONTEXT1_CNTL__DUMMY_PAGE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
75 		MMVM_CONTEXT1_CNTL__PDE0_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
76 		MMVM_CONTEXT1_CNTL__VALID_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
77 		MMVM_CONTEXT1_CNTL__READ_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
78 		MMVM_CONTEXT1_CNTL__WRITE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK |
79 		MMVM_CONTEXT1_CNTL__EXECUTE_PROTECTION_FAULT_ENABLE_INTERRUPT_MASK;
80 
81 	switch (state) {
82 	case AMDGPU_IRQ_STATE_DISABLE:
83 		/* MM HUB */
84 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
85 		for (i = 0; i < 16; i++) {
86 			reg = hub->vm_context0_cntl + i;
87 			tmp = RREG32(reg);
88 			tmp &= ~bits[AMDGPU_MMHUB_0];
89 			WREG32(reg, tmp);
90 		}
91 
92 		/* GFX HUB */
93 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
94 		for (i = 0; i < 16; i++) {
95 			reg = hub->vm_context0_cntl + i;
96 			tmp = RREG32(reg);
97 			tmp &= ~bits[AMDGPU_GFXHUB_0];
98 			WREG32(reg, tmp);
99 		}
100 		break;
101 	case AMDGPU_IRQ_STATE_ENABLE:
102 		/* MM HUB */
103 		hub = &adev->vmhub[AMDGPU_MMHUB_0];
104 		for (i = 0; i < 16; i++) {
105 			reg = hub->vm_context0_cntl + i;
106 			tmp = RREG32(reg);
107 			tmp |= bits[AMDGPU_MMHUB_0];
108 			WREG32(reg, tmp);
109 		}
110 
111 		/* GFX HUB */
112 		hub = &adev->vmhub[AMDGPU_GFXHUB_0];
113 		for (i = 0; i < 16; i++) {
114 			reg = hub->vm_context0_cntl + i;
115 			tmp = RREG32(reg);
116 			tmp |= bits[AMDGPU_GFXHUB_0];
117 			WREG32(reg, tmp);
118 		}
119 		break;
120 	default:
121 		break;
122 	}
123 
124 	return 0;
125 }
126 
127 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev,
128 				       struct amdgpu_irq_src *source,
129 				       struct amdgpu_iv_entry *entry)
130 {
131 	struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
132 	uint32_t status = 0;
133 	u64 addr;
134 
135 	addr = (u64)entry->src_data[0] << 12;
136 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
137 
138 	if (!amdgpu_sriov_vf(adev)) {
139 		/*
140 		 * Issue a dummy read to wait for the status register to
141 		 * be updated to avoid reading an incorrect value due to
142 		 * the new fast GRBM interface.
143 		 */
144 		if (entry->vmid_src == AMDGPU_GFXHUB_0)
145 			RREG32(hub->vm_l2_pro_fault_status);
146 
147 		status = RREG32(hub->vm_l2_pro_fault_status);
148 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
149 	}
150 
151 	if (printk_ratelimit()) {
152 		struct amdgpu_task_info task_info;
153 
154 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
155 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
156 
157 		dev_err(adev->dev,
158 			"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, "
159 			"for process %s pid %d thread %s pid %d)\n",
160 			entry->vmid_src ? "mmhub" : "gfxhub",
161 			entry->src_id, entry->ring_id, entry->vmid,
162 			entry->pasid, task_info.process_name, task_info.tgid,
163 			task_info.task_name, task_info.pid);
164 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
165 			addr, entry->client_id);
166 		if (!amdgpu_sriov_vf(adev)) {
167 			dev_err(adev->dev,
168 				"GCVM_L2_PROTECTION_FAULT_STATUS:0x%08X\n",
169 				status);
170 			dev_err(adev->dev, "\t MORE_FAULTS: 0x%lx\n",
171 				REG_GET_FIELD(status,
172 				GCVM_L2_PROTECTION_FAULT_STATUS, MORE_FAULTS));
173 			dev_err(adev->dev, "\t WALKER_ERROR: 0x%lx\n",
174 				REG_GET_FIELD(status,
175 				GCVM_L2_PROTECTION_FAULT_STATUS, WALKER_ERROR));
176 			dev_err(adev->dev, "\t PERMISSION_FAULTS: 0x%lx\n",
177 				REG_GET_FIELD(status,
178 				GCVM_L2_PROTECTION_FAULT_STATUS, PERMISSION_FAULTS));
179 			dev_err(adev->dev, "\t MAPPING_ERROR: 0x%lx\n",
180 				REG_GET_FIELD(status,
181 				GCVM_L2_PROTECTION_FAULT_STATUS, MAPPING_ERROR));
182 			dev_err(adev->dev, "\t RW: 0x%lx\n",
183 				REG_GET_FIELD(status,
184 				GCVM_L2_PROTECTION_FAULT_STATUS, RW));
185 		}
186 	}
187 
188 	return 0;
189 }
190 
191 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = {
192 	.set = gmc_v10_0_vm_fault_interrupt_state,
193 	.process = gmc_v10_0_process_interrupt,
194 };
195 
196 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev)
197 {
198 	adev->gmc.vm_fault.num_types = 1;
199 	adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs;
200 }
201 
202 static uint32_t gmc_v10_0_get_invalidate_req(unsigned int vmid,
203 					     uint32_t flush_type)
204 {
205 	u32 req = 0;
206 
207 	/* invalidate using legacy mode on vmid*/
208 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ,
209 			    PER_VMID_INVALIDATE_REQ, 1 << vmid);
210 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, FLUSH_TYPE, flush_type);
211 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PTES, 1);
212 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE0, 1);
213 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE1, 1);
214 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L2_PDE2, 1);
215 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ, INVALIDATE_L1_PTES, 1);
216 	req = REG_SET_FIELD(req, GCVM_INVALIDATE_ENG0_REQ,
217 			    CLEAR_PROTECTION_FAULT_STATUS_ADDR,	0);
218 
219 	return req;
220 }
221 
222 /**
223  * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore
224  *
225  * @adev: amdgpu_device pointer
226  * @vmhub: vmhub type
227  *
228  */
229 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev,
230 				       uint32_t vmhub)
231 {
232 	return ((vmhub == AMDGPU_MMHUB_0 ||
233 		 vmhub == AMDGPU_MMHUB_1) &&
234 		(!amdgpu_sriov_vf(adev)));
235 }
236 
237 /*
238  * GART
239  * VMID 0 is the physical GPU addresses as used by the kernel.
240  * VMIDs 1-15 are used for userspace clients and are handled
241  * by the amdgpu vm/hsa code.
242  */
243 
244 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
245 				   unsigned int vmhub, uint32_t flush_type)
246 {
247 	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub);
248 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
249 	u32 tmp = gmc_v10_0_get_invalidate_req(vmid, flush_type);
250 	/* Use register 17 for GART */
251 	const unsigned eng = 17;
252 	unsigned int i;
253 
254 	spin_lock(&adev->gmc.invalidate_lock);
255 	/*
256 	 * It may lose gpuvm invalidate acknowldege state across power-gating
257 	 * off cycle, add semaphore acquire before invalidation and semaphore
258 	 * release after invalidation to avoid entering power gated state
259 	 * to WA the Issue
260 	 */
261 
262 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
263 	if (use_semaphore) {
264 		for (i = 0; i < adev->usec_timeout; i++) {
265 			/* a read return value of 1 means semaphore acuqire */
266 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng);
267 			if (tmp & 0x1)
268 				break;
269 			udelay(1);
270 		}
271 
272 		if (i >= adev->usec_timeout)
273 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
274 	}
275 
276 	WREG32_NO_KIQ(hub->vm_inv_eng0_req + eng, tmp);
277 
278 	/*
279 	 * Issue a dummy read to wait for the ACK register to be cleared
280 	 * to avoid a false ACK due to the new fast GRBM interface.
281 	 */
282 	if (vmhub == AMDGPU_GFXHUB_0)
283 		RREG32_NO_KIQ(hub->vm_inv_eng0_req + eng);
284 
285 	/* Wait for ACK with a delay.*/
286 	for (i = 0; i < adev->usec_timeout; i++) {
287 		tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack + eng);
288 		tmp &= 1 << vmid;
289 		if (tmp)
290 			break;
291 
292 		udelay(1);
293 	}
294 
295 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
296 	if (use_semaphore)
297 		/*
298 		 * add semaphore release after invalidation,
299 		 * write with 0 means semaphore release
300 		 */
301 		WREG32_NO_KIQ(hub->vm_inv_eng0_sem + eng, 0);
302 
303 	spin_unlock(&adev->gmc.invalidate_lock);
304 
305 	if (i < adev->usec_timeout)
306 		return;
307 
308 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
309 }
310 
311 /**
312  * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback
313  *
314  * @adev: amdgpu_device pointer
315  * @vmid: vm instance to flush
316  *
317  * Flush the TLB for the requested page table.
318  */
319 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
320 					uint32_t vmhub, uint32_t flush_type)
321 {
322 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
323 	struct dma_fence *fence;
324 	struct amdgpu_job *job;
325 
326 	int r;
327 
328 	/* flush hdp cache */
329 	adev->nbio.funcs->hdp_flush(adev, NULL);
330 
331 	mutex_lock(&adev->mman.gtt_window_lock);
332 
333 	if (vmhub == AMDGPU_MMHUB_0) {
334 		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0);
335 		mutex_unlock(&adev->mman.gtt_window_lock);
336 		return;
337 	}
338 
339 	BUG_ON(vmhub != AMDGPU_GFXHUB_0);
340 
341 	if (!adev->mman.buffer_funcs_enabled ||
342 	    !adev->ib_pool_ready ||
343 	    adev->in_gpu_reset ||
344 	    ring->sched.ready == false) {
345 		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0);
346 		mutex_unlock(&adev->mman.gtt_window_lock);
347 		return;
348 	}
349 
350 	/* The SDMA on Navi has a bug which can theoretically result in memory
351 	 * corruption if an invalidation happens at the same time as an VA
352 	 * translation. Avoid this by doing the invalidation from the SDMA
353 	 * itself.
354 	 */
355 	r = amdgpu_job_alloc_with_ib(adev, 16 * 4, &job);
356 	if (r)
357 		goto error_alloc;
358 
359 	job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
360 	job->vm_needs_flush = true;
361 	job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
362 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
363 	r = amdgpu_job_submit(job, &adev->mman.entity,
364 			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
365 	if (r)
366 		goto error_submit;
367 
368 	mutex_unlock(&adev->mman.gtt_window_lock);
369 
370 	dma_fence_wait(fence, false);
371 	dma_fence_put(fence);
372 
373 	return;
374 
375 error_submit:
376 	amdgpu_job_free(job);
377 
378 error_alloc:
379 	mutex_unlock(&adev->mman.gtt_window_lock);
380 	DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
381 }
382 
383 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
384 					     unsigned vmid, uint64_t pd_addr)
385 {
386 	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
387 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
388 	uint32_t req = gmc_v10_0_get_invalidate_req(vmid, 0);
389 	unsigned eng = ring->vm_inv_eng;
390 
391 	/*
392 	 * It may lose gpuvm invalidate acknowldege state across power-gating
393 	 * off cycle, add semaphore acquire before invalidation and semaphore
394 	 * release after invalidation to avoid entering power gated state
395 	 * to WA the Issue
396 	 */
397 
398 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
399 	if (use_semaphore)
400 		/* a read return value of 1 means semaphore acuqire */
401 		amdgpu_ring_emit_reg_wait(ring,
402 					  hub->vm_inv_eng0_sem + eng, 0x1, 0x1);
403 
404 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 + (2 * vmid),
405 			      lower_32_bits(pd_addr));
406 
407 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 * vmid),
408 			      upper_32_bits(pd_addr));
409 
410 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req + eng,
411 					    hub->vm_inv_eng0_ack + eng,
412 					    req, 1 << vmid);
413 
414 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
415 	if (use_semaphore)
416 		/*
417 		 * add semaphore release after invalidation,
418 		 * write with 0 means semaphore release
419 		 */
420 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem + eng, 0);
421 
422 	return pd_addr;
423 }
424 
425 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
426 					 unsigned pasid)
427 {
428 	struct amdgpu_device *adev = ring->adev;
429 	uint32_t reg;
430 
431 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
432 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
433 	else
434 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
435 
436 	amdgpu_ring_emit_wreg(ring, reg, pasid);
437 }
438 
439 /*
440  * PTE format on NAVI 10:
441  * 63:59 reserved
442  * 58:57 reserved
443  * 56 F
444  * 55 L
445  * 54 reserved
446  * 53:52 SW
447  * 51 T
448  * 50:48 mtype
449  * 47:12 4k physical page base address
450  * 11:7 fragment
451  * 6 write
452  * 5 read
453  * 4 exe
454  * 3 Z
455  * 2 snooped
456  * 1 system
457  * 0 valid
458  *
459  * PDE format on NAVI 10:
460  * 63:59 block fragment size
461  * 58:55 reserved
462  * 54 P
463  * 53:48 reserved
464  * 47:6 physical base address of PD or PTE
465  * 5:3 reserved
466  * 2 C
467  * 1 system
468  * 0 valid
469  */
470 
471 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
472 {
473 	switch (flags) {
474 	case AMDGPU_VM_MTYPE_DEFAULT:
475 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
476 	case AMDGPU_VM_MTYPE_NC:
477 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
478 	case AMDGPU_VM_MTYPE_WC:
479 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
480 	case AMDGPU_VM_MTYPE_CC:
481 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
482 	case AMDGPU_VM_MTYPE_UC:
483 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
484 	default:
485 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
486 	}
487 }
488 
489 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level,
490 				 uint64_t *addr, uint64_t *flags)
491 {
492 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
493 		*addr = adev->vm_manager.vram_base_offset + *addr -
494 			adev->gmc.vram_start;
495 	BUG_ON(*addr & 0xFFFF00000000003FULL);
496 
497 	if (!adev->gmc.translate_further)
498 		return;
499 
500 	if (level == AMDGPU_VM_PDB1) {
501 		/* Set the block fragment size */
502 		if (!(*flags & AMDGPU_PDE_PTE))
503 			*flags |= AMDGPU_PDE_BFS(0x9);
504 
505 	} else if (level == AMDGPU_VM_PDB0) {
506 		if (*flags & AMDGPU_PDE_PTE)
507 			*flags &= ~AMDGPU_PDE_PTE;
508 		else
509 			*flags |= AMDGPU_PTE_TF;
510 	}
511 }
512 
513 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
514 				 struct amdgpu_bo_va_mapping *mapping,
515 				 uint64_t *flags)
516 {
517 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
518 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
519 
520 	*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
521 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
522 
523 	if (mapping->flags & AMDGPU_PTE_PRT) {
524 		*flags |= AMDGPU_PTE_PRT;
525 		*flags |= AMDGPU_PTE_SNOOPED;
526 		*flags |= AMDGPU_PTE_LOG;
527 		*flags |= AMDGPU_PTE_SYSTEM;
528 		*flags &= ~AMDGPU_PTE_VALID;
529 	}
530 }
531 
532 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = {
533 	.flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb,
534 	.emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb,
535 	.emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping,
536 	.map_mtype = gmc_v10_0_map_mtype,
537 	.get_vm_pde = gmc_v10_0_get_vm_pde,
538 	.get_vm_pte = gmc_v10_0_get_vm_pte
539 };
540 
541 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev)
542 {
543 	if (adev->gmc.gmc_funcs == NULL)
544 		adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs;
545 }
546 
547 static int gmc_v10_0_early_init(void *handle)
548 {
549 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
550 
551 	gmc_v10_0_set_gmc_funcs(adev);
552 	gmc_v10_0_set_irq_funcs(adev);
553 
554 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
555 	adev->gmc.shared_aperture_end =
556 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
557 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
558 	adev->gmc.private_aperture_end =
559 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
560 
561 	return 0;
562 }
563 
564 static int gmc_v10_0_late_init(void *handle)
565 {
566 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
567 	unsigned vm_inv_eng[AMDGPU_MAX_VMHUBS] = { 4, 4 };
568 	unsigned i;
569 
570 	for(i = 0; i < adev->num_rings; ++i) {
571 		struct amdgpu_ring *ring = adev->rings[i];
572 		unsigned vmhub = ring->funcs->vmhub;
573 
574 		ring->vm_inv_eng = vm_inv_eng[vmhub]++;
575 		dev_info(adev->dev, "ring %u(%s) uses VM inv eng %u on hub %u\n",
576 			 ring->idx, ring->name, ring->vm_inv_eng,
577 			 ring->funcs->vmhub);
578 	}
579 
580 	/* Engine 17 is used for GART flushes */
581 	for(i = 0; i < AMDGPU_MAX_VMHUBS; ++i)
582 		BUG_ON(vm_inv_eng[i] > 17);
583 
584 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
585 }
586 
587 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev,
588 					struct amdgpu_gmc *mc)
589 {
590 	u64 base = 0;
591 
592 	base = gfxhub_v2_0_get_fb_location(adev);
593 
594 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
595 	amdgpu_gmc_gart_location(adev, mc);
596 
597 	/* base offset of vram pages */
598 	adev->vm_manager.vram_base_offset = gfxhub_v2_0_get_mc_fb_offset(adev);
599 }
600 
601 /**
602  * gmc_v10_0_mc_init - initialize the memory controller driver params
603  *
604  * @adev: amdgpu_device pointer
605  *
606  * Look up the amount of vram, vram width, and decide how to place
607  * vram and gart within the GPU's physical address space.
608  * Returns 0 for success.
609  */
610 static int gmc_v10_0_mc_init(struct amdgpu_device *adev)
611 {
612 	/* Could aper size report 0 ? */
613 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
614 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
615 
616 	/* size in MB on si */
617 	adev->gmc.mc_vram_size =
618 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
619 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
620 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
621 
622 	/* In case the PCI BAR is larger than the actual amount of vram */
623 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
624 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
625 
626 	/* set the gart size */
627 	if (amdgpu_gart_size == -1) {
628 		switch (adev->asic_type) {
629 		case CHIP_NAVI10:
630 		case CHIP_NAVI14:
631 		case CHIP_NAVI12:
632 		default:
633 			adev->gmc.gart_size = 512ULL << 20;
634 			break;
635 		}
636 	} else
637 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
638 
639 	gmc_v10_0_vram_gtt_location(adev, &adev->gmc);
640 
641 	return 0;
642 }
643 
644 static int gmc_v10_0_gart_init(struct amdgpu_device *adev)
645 {
646 	int r;
647 
648 	if (adev->gart.bo) {
649 		WARN(1, "NAVI10 PCIE GART already initialized\n");
650 		return 0;
651 	}
652 
653 	/* Initialize common gart structure */
654 	r = amdgpu_gart_init(adev);
655 	if (r)
656 		return r;
657 
658 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
659 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
660 				 AMDGPU_PTE_EXECUTABLE;
661 
662 	return amdgpu_gart_table_vram_alloc(adev);
663 }
664 
665 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev)
666 {
667 	u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
668 	unsigned size;
669 
670 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
671 		size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
672 	} else {
673 		u32 viewport;
674 		u32 pitch;
675 
676 		viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
677 		pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH);
678 		size = (REG_GET_FIELD(viewport,
679 					HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
680 				REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) *
681 				4);
682 	}
683 	/* return 0 if the pre-OS buffer uses up most of vram */
684 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024)) {
685 		DRM_ERROR("Warning: pre-OS buffer uses most of vram, \
686 				be aware of gart table overwrite\n");
687 		return 0;
688 	}
689 
690 	return size;
691 }
692 
693 
694 
695 static int gmc_v10_0_sw_init(void *handle)
696 {
697 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
698 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
699 
700 	gfxhub_v2_0_init(adev);
701 	mmhub_v2_0_init(adev);
702 
703 	spin_lock_init(&adev->gmc.invalidate_lock);
704 
705 	r = amdgpu_atomfirmware_get_vram_info(adev,
706 		&vram_width, &vram_type, &vram_vendor);
707 	if (!amdgpu_emu_mode)
708 		adev->gmc.vram_width = vram_width;
709 	else
710 		adev->gmc.vram_width = 1 * 128; /* numchan * chansize */
711 
712 	adev->gmc.vram_type = vram_type;
713 	adev->gmc.vram_vendor = vram_vendor;
714 	switch (adev->asic_type) {
715 	case CHIP_NAVI10:
716 	case CHIP_NAVI14:
717 	case CHIP_NAVI12:
718 		adev->num_vmhubs = 2;
719 		/*
720 		 * To fulfill 4-level page support,
721 		 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12,
722 		 * block size 512 (9bit)
723 		 */
724 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
725 		break;
726 	default:
727 		break;
728 	}
729 
730 	/* This interrupt is VMC page fault.*/
731 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC,
732 			      VMC_1_0__SRCID__VM_FAULT,
733 			      &adev->gmc.vm_fault);
734 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2,
735 			      UTCL2_1_0__SRCID__FAULT,
736 			      &adev->gmc.vm_fault);
737 	if (r)
738 		return r;
739 
740 	/*
741 	 * Set the internal MC address mask This is the max address of the GPU's
742 	 * internal address space.
743 	 */
744 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
745 
746 	/*
747 	 * Reserve 8M stolen memory for navi10 like vega10
748 	 * TODO: will check if it's really needed on asic.
749 	 */
750 	if (amdgpu_emu_mode == 1)
751 		adev->gmc.stolen_size = 0;
752 	else
753 		adev->gmc.stolen_size = 9 * 1024 *1024;
754 
755 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
756 	if (r) {
757 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
758 		return r;
759 	}
760 
761 	r = gmc_v10_0_mc_init(adev);
762 	if (r)
763 		return r;
764 
765 	adev->gmc.stolen_size = gmc_v10_0_get_vbios_fb_size(adev);
766 
767 	/* Memory manager */
768 	r = amdgpu_bo_init(adev);
769 	if (r)
770 		return r;
771 
772 	r = gmc_v10_0_gart_init(adev);
773 	if (r)
774 		return r;
775 
776 	/*
777 	 * number of VMs
778 	 * VMID 0 is reserved for System
779 	 * amdgpu graphics/compute will use VMIDs 1-7
780 	 * amdkfd will use VMIDs 8-15
781 	 */
782 	adev->vm_manager.id_mgr[AMDGPU_GFXHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
783 	adev->vm_manager.id_mgr[AMDGPU_MMHUB_0].num_ids = AMDGPU_NUM_OF_VMIDS;
784 
785 	amdgpu_vm_manager_init(adev);
786 
787 	return 0;
788 }
789 
790 /**
791  * gmc_v8_0_gart_fini - vm fini callback
792  *
793  * @adev: amdgpu_device pointer
794  *
795  * Tears down the driver GART/VM setup (CIK).
796  */
797 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev)
798 {
799 	amdgpu_gart_table_vram_free(adev);
800 	amdgpu_gart_fini(adev);
801 }
802 
803 static int gmc_v10_0_sw_fini(void *handle)
804 {
805 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
806 
807 	amdgpu_vm_manager_fini(adev);
808 	gmc_v10_0_gart_fini(adev);
809 	amdgpu_gem_force_release(adev);
810 	amdgpu_bo_fini(adev);
811 
812 	return 0;
813 }
814 
815 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev)
816 {
817 	switch (adev->asic_type) {
818 	case CHIP_NAVI10:
819 	case CHIP_NAVI14:
820 	case CHIP_NAVI12:
821 		break;
822 	default:
823 		break;
824 	}
825 }
826 
827 /**
828  * gmc_v10_0_gart_enable - gart enable
829  *
830  * @adev: amdgpu_device pointer
831  */
832 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev)
833 {
834 	int r;
835 	bool value;
836 	u32 tmp;
837 
838 	if (adev->gart.bo == NULL) {
839 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
840 		return -EINVAL;
841 	}
842 
843 	r = amdgpu_gart_table_vram_pin(adev);
844 	if (r)
845 		return r;
846 
847 	r = gfxhub_v2_0_gart_enable(adev);
848 	if (r)
849 		return r;
850 
851 	r = mmhub_v2_0_gart_enable(adev);
852 	if (r)
853 		return r;
854 
855 	tmp = RREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL);
856 	tmp |= HDP_MISC_CNTL__FLUSH_INVALIDATE_CACHE_MASK;
857 	WREG32_SOC15(HDP, 0, mmHDP_MISC_CNTL, tmp);
858 
859 	tmp = RREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL);
860 	WREG32_SOC15(HDP, 0, mmHDP_HOST_PATH_CNTL, tmp);
861 
862 	/* Flush HDP after it is initialized */
863 	adev->nbio.funcs->hdp_flush(adev, NULL);
864 
865 	value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
866 		false : true;
867 
868 	gfxhub_v2_0_set_fault_enable_default(adev, value);
869 	mmhub_v2_0_set_fault_enable_default(adev, value);
870 	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0);
871 	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0);
872 
873 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
874 		 (unsigned)(adev->gmc.gart_size >> 20),
875 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
876 
877 	adev->gart.ready = true;
878 
879 	return 0;
880 }
881 
882 static int gmc_v10_0_hw_init(void *handle)
883 {
884 	int r;
885 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
886 
887 	/* The sequence of these two function calls matters.*/
888 	gmc_v10_0_init_golden_registers(adev);
889 
890 	r = gmc_v10_0_gart_enable(adev);
891 	if (r)
892 		return r;
893 
894 	return 0;
895 }
896 
897 /**
898  * gmc_v10_0_gart_disable - gart disable
899  *
900  * @adev: amdgpu_device pointer
901  *
902  * This disables all VM page table.
903  */
904 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev)
905 {
906 	gfxhub_v2_0_gart_disable(adev);
907 	mmhub_v2_0_gart_disable(adev);
908 	amdgpu_gart_table_vram_unpin(adev);
909 }
910 
911 static int gmc_v10_0_hw_fini(void *handle)
912 {
913 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
914 
915 	if (amdgpu_sriov_vf(adev)) {
916 		/* full access mode, so don't touch any GMC register */
917 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
918 		return 0;
919 	}
920 
921 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
922 	gmc_v10_0_gart_disable(adev);
923 
924 	return 0;
925 }
926 
927 static int gmc_v10_0_suspend(void *handle)
928 {
929 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
930 
931 	gmc_v10_0_hw_fini(adev);
932 
933 	return 0;
934 }
935 
936 static int gmc_v10_0_resume(void *handle)
937 {
938 	int r;
939 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
940 
941 	r = gmc_v10_0_hw_init(adev);
942 	if (r)
943 		return r;
944 
945 	amdgpu_vmid_reset_all(adev);
946 
947 	return 0;
948 }
949 
950 static bool gmc_v10_0_is_idle(void *handle)
951 {
952 	/* MC is always ready in GMC v10.*/
953 	return true;
954 }
955 
956 static int gmc_v10_0_wait_for_idle(void *handle)
957 {
958 	/* There is no need to wait for MC idle in GMC v10.*/
959 	return 0;
960 }
961 
962 static int gmc_v10_0_soft_reset(void *handle)
963 {
964 	return 0;
965 }
966 
967 static int gmc_v10_0_set_clockgating_state(void *handle,
968 					   enum amd_clockgating_state state)
969 {
970 	int r;
971 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
972 
973 	r = mmhub_v2_0_set_clockgating(adev, state);
974 	if (r)
975 		return r;
976 
977 	return athub_v2_0_set_clockgating(adev, state);
978 }
979 
980 static void gmc_v10_0_get_clockgating_state(void *handle, u32 *flags)
981 {
982 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
983 
984 	mmhub_v2_0_get_clockgating(adev, flags);
985 
986 	athub_v2_0_get_clockgating(adev, flags);
987 }
988 
989 static int gmc_v10_0_set_powergating_state(void *handle,
990 					   enum amd_powergating_state state)
991 {
992 	return 0;
993 }
994 
995 const struct amd_ip_funcs gmc_v10_0_ip_funcs = {
996 	.name = "gmc_v10_0",
997 	.early_init = gmc_v10_0_early_init,
998 	.late_init = gmc_v10_0_late_init,
999 	.sw_init = gmc_v10_0_sw_init,
1000 	.sw_fini = gmc_v10_0_sw_fini,
1001 	.hw_init = gmc_v10_0_hw_init,
1002 	.hw_fini = gmc_v10_0_hw_fini,
1003 	.suspend = gmc_v10_0_suspend,
1004 	.resume = gmc_v10_0_resume,
1005 	.is_idle = gmc_v10_0_is_idle,
1006 	.wait_for_idle = gmc_v10_0_wait_for_idle,
1007 	.soft_reset = gmc_v10_0_soft_reset,
1008 	.set_clockgating_state = gmc_v10_0_set_clockgating_state,
1009 	.set_powergating_state = gmc_v10_0_set_powergating_state,
1010 	.get_clockgating_state = gmc_v10_0_get_clockgating_state,
1011 };
1012 
1013 const struct amdgpu_ip_block_version gmc_v10_0_ip_block =
1014 {
1015 	.type = AMD_IP_BLOCK_TYPE_GMC,
1016 	.major = 10,
1017 	.minor = 0,
1018 	.rev = 0,
1019 	.funcs = &gmc_v10_0_ip_funcs,
1020 };
1021