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 
26 #include <drm/drm_cache.h>
27 
28 #include "amdgpu.h"
29 #include "amdgpu_atomfirmware.h"
30 #include "gmc_v10_0.h"
31 #include "umc_v8_7.h"
32 
33 #include "athub/athub_2_0_0_sh_mask.h"
34 #include "athub/athub_2_0_0_offset.h"
35 #include "dcn/dcn_2_0_0_offset.h"
36 #include "dcn/dcn_2_0_0_sh_mask.h"
37 #include "oss/osssys_5_0_0_offset.h"
38 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
39 #include "navi10_enum.h"
40 
41 #include "soc15.h"
42 #include "soc15d.h"
43 #include "soc15_common.h"
44 
45 #include "nbio_v2_3.h"
46 
47 #include "gfxhub_v2_0.h"
48 #include "gfxhub_v2_1.h"
49 #include "mmhub_v2_0.h"
50 #include "mmhub_v2_3.h"
51 #include "athub_v2_0.h"
52 #include "athub_v2_1.h"
53 
54 #include "amdgpu_reset.h"
55 
56 #if 0
57 static const struct soc15_reg_golden golden_settings_navi10_hdp[] =
58 {
59 	/* TODO add golden setting for hdp */
60 };
61 #endif
62 
63 static int gmc_v10_0_ecc_interrupt_state(struct amdgpu_device *adev,
64 					 struct amdgpu_irq_src *src,
65 					 unsigned type,
66 					 enum amdgpu_interrupt_state state)
67 {
68 	return 0;
69 }
70 
71 static int
72 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
73 				   struct amdgpu_irq_src *src, unsigned type,
74 				   enum amdgpu_interrupt_state state)
75 {
76 	switch (state) {
77 	case AMDGPU_IRQ_STATE_DISABLE:
78 		/* MM HUB */
79 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, false);
80 		/* GFX HUB */
81 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, false);
82 		break;
83 	case AMDGPU_IRQ_STATE_ENABLE:
84 		/* MM HUB */
85 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, true);
86 		/* GFX HUB */
87 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, true);
88 		break;
89 	default:
90 		break;
91 	}
92 
93 	return 0;
94 }
95 
96 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev,
97 				       struct amdgpu_irq_src *source,
98 				       struct amdgpu_iv_entry *entry)
99 {
100 	bool retry_fault = !!(entry->src_data[1] & 0x80);
101 	bool write_fault = !!(entry->src_data[1] & 0x20);
102 	struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
103 	struct amdgpu_task_info task_info;
104 	uint32_t status = 0;
105 	u64 addr;
106 
107 	addr = (u64)entry->src_data[0] << 12;
108 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
109 
110 	if (retry_fault) {
111 		/* Returning 1 here also prevents sending the IV to the KFD */
112 
113 		/* Process it onyl if it's the first fault for this address */
114 		if (entry->ih != &adev->irq.ih_soft &&
115 		    amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid,
116 					     entry->timestamp))
117 			return 1;
118 
119 		/* Delegate it to a different ring if the hardware hasn't
120 		 * already done it.
121 		 */
122 		if (entry->ih == &adev->irq.ih) {
123 			amdgpu_irq_delegate(adev, entry, 8);
124 			return 1;
125 		}
126 
127 		/* Try to handle the recoverable page faults by filling page
128 		 * tables
129 		 */
130 		if (amdgpu_vm_handle_fault(adev, entry->pasid, addr, write_fault))
131 			return 1;
132 	}
133 
134 	if (!amdgpu_sriov_vf(adev)) {
135 		/*
136 		 * Issue a dummy read to wait for the status register to
137 		 * be updated to avoid reading an incorrect value due to
138 		 * the new fast GRBM interface.
139 		 */
140 		if ((entry->vmid_src == AMDGPU_GFXHUB_0) &&
141 		    (adev->ip_versions[GC_HWIP][0] < IP_VERSION(10, 3, 0)))
142 			RREG32(hub->vm_l2_pro_fault_status);
143 
144 		status = RREG32(hub->vm_l2_pro_fault_status);
145 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
146 	}
147 
148 	if (!printk_ratelimit())
149 		return 0;
150 
151 	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
152 	amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
153 
154 	dev_err(adev->dev,
155 		"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, "
156 		"for process %s pid %d thread %s pid %d)\n",
157 		entry->vmid_src ? "mmhub" : "gfxhub",
158 		entry->src_id, entry->ring_id, entry->vmid,
159 		entry->pasid, task_info.process_name, task_info.tgid,
160 		task_info.task_name, task_info.pid);
161 	dev_err(adev->dev, "  in page starting at address 0x%016llx from client 0x%x (%s)\n",
162 		addr, entry->client_id,
163 		soc15_ih_clientid_name[entry->client_id]);
164 
165 	if (!amdgpu_sriov_vf(adev))
166 		hub->vmhub_funcs->print_l2_protection_fault_status(adev,
167 								   status);
168 
169 	return 0;
170 }
171 
172 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = {
173 	.set = gmc_v10_0_vm_fault_interrupt_state,
174 	.process = gmc_v10_0_process_interrupt,
175 };
176 
177 static const struct amdgpu_irq_src_funcs gmc_v10_0_ecc_funcs = {
178 	.set = gmc_v10_0_ecc_interrupt_state,
179 	.process = amdgpu_umc_process_ecc_irq,
180 };
181 
182 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev)
183 {
184 	adev->gmc.vm_fault.num_types = 1;
185 	adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs;
186 
187 	if (!amdgpu_sriov_vf(adev)) {
188 		adev->gmc.ecc_irq.num_types = 1;
189 		adev->gmc.ecc_irq.funcs = &gmc_v10_0_ecc_funcs;
190 	}
191 }
192 
193 /**
194  * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore
195  *
196  * @adev: amdgpu_device pointer
197  * @vmhub: vmhub type
198  *
199  */
200 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev,
201 				       uint32_t vmhub)
202 {
203 	return ((vmhub == AMDGPU_MMHUB_0 ||
204 		 vmhub == AMDGPU_MMHUB_1) &&
205 		(!amdgpu_sriov_vf(adev)));
206 }
207 
208 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info(
209 					struct amdgpu_device *adev,
210 					uint8_t vmid, uint16_t *p_pasid)
211 {
212 	uint32_t value;
213 
214 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
215 		     + vmid);
216 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
217 
218 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
219 }
220 
221 /*
222  * GART
223  * VMID 0 is the physical GPU addresses as used by the kernel.
224  * VMIDs 1-15 are used for userspace clients and are handled
225  * by the amdgpu vm/hsa code.
226  */
227 
228 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
229 				   unsigned int vmhub, uint32_t flush_type)
230 {
231 	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub);
232 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
233 	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
234 	u32 tmp;
235 	/* Use register 17 for GART */
236 	const unsigned eng = 17;
237 	unsigned int i;
238 	unsigned char hub_ip = 0;
239 
240 	hub_ip = (vmhub == AMDGPU_GFXHUB_0) ?
241 		   GC_HWIP : MMHUB_HWIP;
242 
243 	spin_lock(&adev->gmc.invalidate_lock);
244 	/*
245 	 * It may lose gpuvm invalidate acknowldege state across power-gating
246 	 * off cycle, add semaphore acquire before invalidation and semaphore
247 	 * release after invalidation to avoid entering power gated state
248 	 * to WA the Issue
249 	 */
250 
251 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
252 	if (use_semaphore) {
253 		for (i = 0; i < adev->usec_timeout; i++) {
254 			/* a read return value of 1 means semaphore acuqire */
255 			tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
256 					 hub->eng_distance * eng, hub_ip);
257 
258 			if (tmp & 0x1)
259 				break;
260 			udelay(1);
261 		}
262 
263 		if (i >= adev->usec_timeout)
264 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
265 	}
266 
267 	WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req +
268 			  hub->eng_distance * eng,
269 			  inv_req, hub_ip);
270 
271 	/*
272 	 * Issue a dummy read to wait for the ACK register to be cleared
273 	 * to avoid a false ACK due to the new fast GRBM interface.
274 	 */
275 	if ((vmhub == AMDGPU_GFXHUB_0) &&
276 	    (adev->ip_versions[GC_HWIP][0] < IP_VERSION(10, 3, 0)))
277 		RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req +
278 				  hub->eng_distance * eng, hub_ip);
279 
280 	/* Wait for ACK with a delay.*/
281 	for (i = 0; i < adev->usec_timeout; i++) {
282 		tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack +
283 				  hub->eng_distance * eng, hub_ip);
284 
285 		tmp &= 1 << vmid;
286 		if (tmp)
287 			break;
288 
289 		udelay(1);
290 	}
291 
292 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
293 	if (use_semaphore)
294 		/*
295 		 * add semaphore release after invalidation,
296 		 * write with 0 means semaphore release
297 		 */
298 		WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
299 				  hub->eng_distance * eng, 0, hub_ip);
300 
301 	spin_unlock(&adev->gmc.invalidate_lock);
302 
303 	if (i < adev->usec_timeout)
304 		return;
305 
306 	DRM_ERROR("Timeout waiting for VM flush hub: %d!\n", vmhub);
307 }
308 
309 /**
310  * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback
311  *
312  * @adev: amdgpu_device pointer
313  * @vmid: vm instance to flush
314  * @vmhub: vmhub type
315  * @flush_type: the flush type
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->hdp.funcs->flush_hdp(adev, NULL);
330 
331 	/* For SRIOV run time, driver shouldn't access the register through MMIO
332 	 * Directly use kiq to do the vm invalidation instead
333 	 */
334 	if (adev->gfx.kiq.ring.sched.ready && !adev->enable_mes &&
335 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
336 	    down_read_trylock(&adev->reset_domain->sem)) {
337 		struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
338 		const unsigned eng = 17;
339 		u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
340 		u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
341 		u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
342 
343 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
344 				1 << vmid);
345 
346 		up_read(&adev->reset_domain->sem);
347 		return;
348 	}
349 
350 	mutex_lock(&adev->mman.gtt_window_lock);
351 
352 	if (vmhub == AMDGPU_MMHUB_0) {
353 		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB_0, 0);
354 		mutex_unlock(&adev->mman.gtt_window_lock);
355 		return;
356 	}
357 
358 	BUG_ON(vmhub != AMDGPU_GFXHUB_0);
359 
360 	if (!adev->mman.buffer_funcs_enabled ||
361 	    !adev->ib_pool_ready ||
362 	    amdgpu_in_reset(adev) ||
363 	    ring->sched.ready == false) {
364 		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB_0, 0);
365 		mutex_unlock(&adev->mman.gtt_window_lock);
366 		return;
367 	}
368 
369 	/* The SDMA on Navi has a bug which can theoretically result in memory
370 	 * corruption if an invalidation happens at the same time as an VA
371 	 * translation. Avoid this by doing the invalidation from the SDMA
372 	 * itself.
373 	 */
374 	r = amdgpu_job_alloc_with_ib(adev, 16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
375 				     &job);
376 	if (r)
377 		goto error_alloc;
378 
379 	job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
380 	job->vm_needs_flush = true;
381 	job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
382 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
383 	r = amdgpu_job_submit(job, &adev->mman.entity,
384 			      AMDGPU_FENCE_OWNER_UNDEFINED, &fence);
385 	if (r)
386 		goto error_submit;
387 
388 	mutex_unlock(&adev->mman.gtt_window_lock);
389 
390 	dma_fence_wait(fence, false);
391 	dma_fence_put(fence);
392 
393 	return;
394 
395 error_submit:
396 	amdgpu_job_free(job);
397 
398 error_alloc:
399 	mutex_unlock(&adev->mman.gtt_window_lock);
400 	DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
401 }
402 
403 /**
404  * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid
405  *
406  * @adev: amdgpu_device pointer
407  * @pasid: pasid to be flush
408  * @flush_type: the flush type
409  * @all_hub: Used with PACKET3_INVALIDATE_TLBS_ALL_HUB()
410  *
411  * Flush the TLB for the requested pasid.
412  */
413 static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
414 					uint16_t pasid, uint32_t flush_type,
415 					bool all_hub)
416 {
417 	int vmid, i;
418 	signed long r;
419 	uint32_t seq;
420 	uint16_t queried_pasid;
421 	bool ret;
422 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
423 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
424 
425 	if (amdgpu_emu_mode == 0 && ring->sched.ready) {
426 		spin_lock(&adev->gfx.kiq.ring_lock);
427 		/* 2 dwords flush + 8 dwords fence */
428 		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8);
429 		kiq->pmf->kiq_invalidate_tlbs(ring,
430 					pasid, flush_type, all_hub);
431 		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
432 		if (r) {
433 			amdgpu_ring_undo(ring);
434 			spin_unlock(&adev->gfx.kiq.ring_lock);
435 			return -ETIME;
436 		}
437 
438 		amdgpu_ring_commit(ring);
439 		spin_unlock(&adev->gfx.kiq.ring_lock);
440 		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
441 		if (r < 1) {
442 			dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
443 			return -ETIME;
444 		}
445 
446 		return 0;
447 	}
448 
449 	for (vmid = 1; vmid < AMDGPU_NUM_VMID; vmid++) {
450 
451 		ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
452 				&queried_pasid);
453 		if (ret	&& queried_pasid == pasid) {
454 			if (all_hub) {
455 				for (i = 0; i < adev->num_vmhubs; i++)
456 					gmc_v10_0_flush_gpu_tlb(adev, vmid,
457 							i, flush_type);
458 			} else {
459 				gmc_v10_0_flush_gpu_tlb(adev, vmid,
460 						AMDGPU_GFXHUB_0, flush_type);
461 			}
462 			if (!adev->enable_mes)
463 				break;
464 		}
465 	}
466 
467 	return 0;
468 }
469 
470 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
471 					     unsigned vmid, uint64_t pd_addr)
472 {
473 	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
474 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
475 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
476 	unsigned eng = ring->vm_inv_eng;
477 
478 	/*
479 	 * It may lose gpuvm invalidate acknowldege state across power-gating
480 	 * off cycle, add semaphore acquire before invalidation and semaphore
481 	 * release after invalidation to avoid entering power gated state
482 	 * to WA the Issue
483 	 */
484 
485 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
486 	if (use_semaphore)
487 		/* a read return value of 1 means semaphore acuqire */
488 		amdgpu_ring_emit_reg_wait(ring,
489 					  hub->vm_inv_eng0_sem +
490 					  hub->eng_distance * eng, 0x1, 0x1);
491 
492 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
493 			      (hub->ctx_addr_distance * vmid),
494 			      lower_32_bits(pd_addr));
495 
496 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
497 			      (hub->ctx_addr_distance * vmid),
498 			      upper_32_bits(pd_addr));
499 
500 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
501 					    hub->eng_distance * eng,
502 					    hub->vm_inv_eng0_ack +
503 					    hub->eng_distance * eng,
504 					    req, 1 << vmid);
505 
506 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
507 	if (use_semaphore)
508 		/*
509 		 * add semaphore release after invalidation,
510 		 * write with 0 means semaphore release
511 		 */
512 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
513 				      hub->eng_distance * eng, 0);
514 
515 	return pd_addr;
516 }
517 
518 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
519 					 unsigned pasid)
520 {
521 	struct amdgpu_device *adev = ring->adev;
522 	uint32_t reg;
523 
524 	/* MES fw manages IH_VMID_x_LUT updating */
525 	if (ring->is_mes_queue)
526 		return;
527 
528 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
529 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
530 	else
531 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
532 
533 	amdgpu_ring_emit_wreg(ring, reg, pasid);
534 }
535 
536 /*
537  * PTE format on NAVI 10:
538  * 63:59 reserved
539  * 58 reserved and for sienna_cichlid is used for MALL noalloc
540  * 57 reserved
541  * 56 F
542  * 55 L
543  * 54 reserved
544  * 53:52 SW
545  * 51 T
546  * 50:48 mtype
547  * 47:12 4k physical page base address
548  * 11:7 fragment
549  * 6 write
550  * 5 read
551  * 4 exe
552  * 3 Z
553  * 2 snooped
554  * 1 system
555  * 0 valid
556  *
557  * PDE format on NAVI 10:
558  * 63:59 block fragment size
559  * 58:55 reserved
560  * 54 P
561  * 53:48 reserved
562  * 47:6 physical base address of PD or PTE
563  * 5:3 reserved
564  * 2 C
565  * 1 system
566  * 0 valid
567  */
568 
569 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
570 {
571 	switch (flags) {
572 	case AMDGPU_VM_MTYPE_DEFAULT:
573 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
574 	case AMDGPU_VM_MTYPE_NC:
575 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
576 	case AMDGPU_VM_MTYPE_WC:
577 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
578 	case AMDGPU_VM_MTYPE_CC:
579 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
580 	case AMDGPU_VM_MTYPE_UC:
581 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
582 	default:
583 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
584 	}
585 }
586 
587 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level,
588 				 uint64_t *addr, uint64_t *flags)
589 {
590 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
591 		*addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
592 	BUG_ON(*addr & 0xFFFF00000000003FULL);
593 
594 	if (!adev->gmc.translate_further)
595 		return;
596 
597 	if (level == AMDGPU_VM_PDB1) {
598 		/* Set the block fragment size */
599 		if (!(*flags & AMDGPU_PDE_PTE))
600 			*flags |= AMDGPU_PDE_BFS(0x9);
601 
602 	} else if (level == AMDGPU_VM_PDB0) {
603 		if (*flags & AMDGPU_PDE_PTE)
604 			*flags &= ~AMDGPU_PDE_PTE;
605 		else
606 			*flags |= AMDGPU_PTE_TF;
607 	}
608 }
609 
610 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
611 				 struct amdgpu_bo_va_mapping *mapping,
612 				 uint64_t *flags)
613 {
614 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
615 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
616 
617 	*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
618 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
619 
620 	*flags &= ~AMDGPU_PTE_NOALLOC;
621 	*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
622 
623 	if (mapping->flags & AMDGPU_PTE_PRT) {
624 		*flags |= AMDGPU_PTE_PRT;
625 		*flags |= AMDGPU_PTE_SNOOPED;
626 		*flags |= AMDGPU_PTE_LOG;
627 		*flags |= AMDGPU_PTE_SYSTEM;
628 		*flags &= ~AMDGPU_PTE_VALID;
629 	}
630 }
631 
632 static unsigned gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev)
633 {
634 	u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
635 	unsigned size;
636 
637 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
638 		size = AMDGPU_VBIOS_VGA_ALLOCATION;
639 	} else {
640 		u32 viewport;
641 		u32 pitch;
642 
643 		viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
644 		pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH);
645 		size = (REG_GET_FIELD(viewport,
646 					HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
647 				REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) *
648 				4);
649 	}
650 
651 	return size;
652 }
653 
654 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = {
655 	.flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb,
656 	.flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid,
657 	.emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb,
658 	.emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping,
659 	.map_mtype = gmc_v10_0_map_mtype,
660 	.get_vm_pde = gmc_v10_0_get_vm_pde,
661 	.get_vm_pte = gmc_v10_0_get_vm_pte,
662 	.get_vbios_fb_size = gmc_v10_0_get_vbios_fb_size,
663 };
664 
665 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev)
666 {
667 	if (adev->gmc.gmc_funcs == NULL)
668 		adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs;
669 }
670 
671 static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev)
672 {
673 	switch (adev->ip_versions[UMC_HWIP][0]) {
674 	case IP_VERSION(8, 7, 0):
675 		adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM;
676 		adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM;
677 		adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM;
678 		adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA;
679 		adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0];
680 		adev->umc.ras = &umc_v8_7_ras;
681 		break;
682 	default:
683 		break;
684 	}
685 	if (adev->umc.ras) {
686 		amdgpu_ras_register_ras_block(adev, &adev->umc.ras->ras_block);
687 
688 		strcpy(adev->umc.ras->ras_block.ras_comm.name, "umc");
689 		adev->umc.ras->ras_block.ras_comm.block = AMDGPU_RAS_BLOCK__UMC;
690 		adev->umc.ras->ras_block.ras_comm.type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE;
691 		adev->umc.ras_if = &adev->umc.ras->ras_block.ras_comm;
692 
693 		/* If don't define special ras_late_init function, use default ras_late_init */
694 		if (!adev->umc.ras->ras_block.ras_late_init)
695 				adev->umc.ras->ras_block.ras_late_init = amdgpu_umc_ras_late_init;
696 
697 		/* If not defined special ras_cb function, use default ras_cb */
698 		if (!adev->umc.ras->ras_block.ras_cb)
699 			adev->umc.ras->ras_block.ras_cb = amdgpu_umc_process_ras_data_cb;
700 	}
701 }
702 
703 
704 static void gmc_v10_0_set_mmhub_funcs(struct amdgpu_device *adev)
705 {
706 	switch (adev->ip_versions[MMHUB_HWIP][0]) {
707 	case IP_VERSION(2, 3, 0):
708 	case IP_VERSION(2, 4, 0):
709 	case IP_VERSION(2, 4, 1):
710 		adev->mmhub.funcs = &mmhub_v2_3_funcs;
711 		break;
712 	default:
713 		adev->mmhub.funcs = &mmhub_v2_0_funcs;
714 		break;
715 	}
716 }
717 
718 static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev)
719 {
720 	switch (adev->ip_versions[GC_HWIP][0]) {
721 	case IP_VERSION(10, 3, 0):
722 	case IP_VERSION(10, 3, 2):
723 	case IP_VERSION(10, 3, 1):
724 	case IP_VERSION(10, 3, 4):
725 	case IP_VERSION(10, 3, 5):
726 	case IP_VERSION(10, 3, 6):
727 	case IP_VERSION(10, 3, 3):
728 	case IP_VERSION(10, 3, 7):
729 		adev->gfxhub.funcs = &gfxhub_v2_1_funcs;
730 		break;
731 	default:
732 		adev->gfxhub.funcs = &gfxhub_v2_0_funcs;
733 		break;
734 	}
735 }
736 
737 
738 static int gmc_v10_0_early_init(void *handle)
739 {
740 	int r;
741 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
742 
743 	gmc_v10_0_set_mmhub_funcs(adev);
744 	gmc_v10_0_set_gfxhub_funcs(adev);
745 	gmc_v10_0_set_gmc_funcs(adev);
746 	gmc_v10_0_set_irq_funcs(adev);
747 	gmc_v10_0_set_umc_funcs(adev);
748 
749 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
750 	adev->gmc.shared_aperture_end =
751 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
752 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
753 	adev->gmc.private_aperture_end =
754 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
755 
756 	r = amdgpu_gmc_ras_early_init(adev);
757 	if (r)
758 		return r;
759 
760 	return 0;
761 }
762 
763 static int gmc_v10_0_late_init(void *handle)
764 {
765 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
766 	int r;
767 
768 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
769 	if (r)
770 		return r;
771 
772 	r = amdgpu_gmc_ras_late_init(adev);
773 	if (r)
774 		return r;
775 
776 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
777 }
778 
779 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev,
780 					struct amdgpu_gmc *mc)
781 {
782 	u64 base = 0;
783 
784 	base = adev->gfxhub.funcs->get_fb_location(adev);
785 
786 	/* add the xgmi offset of the physical node */
787 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
788 
789 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
790 	amdgpu_gmc_gart_location(adev, mc);
791 	amdgpu_gmc_agp_location(adev, mc);
792 
793 	/* base offset of vram pages */
794 	adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
795 
796 	/* add the xgmi offset of the physical node */
797 	adev->vm_manager.vram_base_offset +=
798 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
799 }
800 
801 /**
802  * gmc_v10_0_mc_init - initialize the memory controller driver params
803  *
804  * @adev: amdgpu_device pointer
805  *
806  * Look up the amount of vram, vram width, and decide how to place
807  * vram and gart within the GPU's physical address space.
808  * Returns 0 for success.
809  */
810 static int gmc_v10_0_mc_init(struct amdgpu_device *adev)
811 {
812 	int r;
813 
814 	/* size in MB on si */
815 	adev->gmc.mc_vram_size =
816 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
817 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
818 
819 	if (!(adev->flags & AMD_IS_APU)) {
820 		r = amdgpu_device_resize_fb_bar(adev);
821 		if (r)
822 			return r;
823 	}
824 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
825 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
826 
827 #ifdef CONFIG_X86_64
828 	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) {
829 		adev->gmc.aper_base = adev->gfxhub.funcs->get_mc_fb_offset(adev);
830 		adev->gmc.aper_size = adev->gmc.real_vram_size;
831 	}
832 #endif
833 
834 	/* In case the PCI BAR is larger than the actual amount of vram */
835 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
836 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
837 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
838 
839 	/* set the gart size */
840 	if (amdgpu_gart_size == -1) {
841 		switch (adev->ip_versions[GC_HWIP][0]) {
842 		default:
843 			adev->gmc.gart_size = 512ULL << 20;
844 			break;
845 		case IP_VERSION(10, 3, 1):   /* DCE SG support */
846 		case IP_VERSION(10, 3, 3):   /* DCE SG support */
847 		case IP_VERSION(10, 3, 6):   /* DCE SG support */
848 		case IP_VERSION(10, 3, 7):   /* DCE SG support */
849 			adev->gmc.gart_size = 1024ULL << 20;
850 			break;
851 		}
852 	} else {
853 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
854 	}
855 
856 	gmc_v10_0_vram_gtt_location(adev, &adev->gmc);
857 
858 	return 0;
859 }
860 
861 static int gmc_v10_0_gart_init(struct amdgpu_device *adev)
862 {
863 	int r;
864 
865 	if (adev->gart.bo) {
866 		WARN(1, "NAVI10 PCIE GART already initialized\n");
867 		return 0;
868 	}
869 
870 	/* Initialize common gart structure */
871 	r = amdgpu_gart_init(adev);
872 	if (r)
873 		return r;
874 
875 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
876 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
877 				 AMDGPU_PTE_EXECUTABLE;
878 
879 	return amdgpu_gart_table_vram_alloc(adev);
880 }
881 
882 static int gmc_v10_0_sw_init(void *handle)
883 {
884 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
885 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
886 
887 	adev->gfxhub.funcs->init(adev);
888 
889 	adev->mmhub.funcs->init(adev);
890 
891 	spin_lock_init(&adev->gmc.invalidate_lock);
892 
893 	if ((adev->flags & AMD_IS_APU) && amdgpu_emu_mode == 1) {
894 		adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4;
895 		adev->gmc.vram_width = 64;
896 	} else if (amdgpu_emu_mode == 1) {
897 		adev->gmc.vram_type = AMDGPU_VRAM_TYPE_GDDR6;
898 		adev->gmc.vram_width = 1 * 128; /* numchan * chansize */
899 	} else {
900 		r = amdgpu_atomfirmware_get_vram_info(adev,
901 				&vram_width, &vram_type, &vram_vendor);
902 		adev->gmc.vram_width = vram_width;
903 
904 		adev->gmc.vram_type = vram_type;
905 		adev->gmc.vram_vendor = vram_vendor;
906 	}
907 
908 	switch (adev->ip_versions[GC_HWIP][0]) {
909 	case IP_VERSION(10, 3, 0):
910 		adev->gmc.mall_size = 128 * 1024 * 1024;
911 		break;
912 	case IP_VERSION(10, 3, 2):
913 		adev->gmc.mall_size = 96 * 1024 * 1024;
914 		break;
915 	case IP_VERSION(10, 3, 4):
916 		adev->gmc.mall_size = 32 * 1024 * 1024;
917 		break;
918 	case IP_VERSION(10, 3, 5):
919 		adev->gmc.mall_size = 16 * 1024 * 1024;
920 		break;
921 	default:
922 		adev->gmc.mall_size = 0;
923 		break;
924 	}
925 
926 	switch (adev->ip_versions[GC_HWIP][0]) {
927 	case IP_VERSION(10, 1, 10):
928 	case IP_VERSION(10, 1, 1):
929 	case IP_VERSION(10, 1, 2):
930 	case IP_VERSION(10, 1, 3):
931 	case IP_VERSION(10, 1, 4):
932 	case IP_VERSION(10, 3, 0):
933 	case IP_VERSION(10, 3, 2):
934 	case IP_VERSION(10, 3, 1):
935 	case IP_VERSION(10, 3, 4):
936 	case IP_VERSION(10, 3, 5):
937 	case IP_VERSION(10, 3, 6):
938 	case IP_VERSION(10, 3, 3):
939 	case IP_VERSION(10, 3, 7):
940 		adev->num_vmhubs = 2;
941 		/*
942 		 * To fulfill 4-level page support,
943 		 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12,
944 		 * block size 512 (9bit)
945 		 */
946 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
947 		break;
948 	default:
949 		break;
950 	}
951 
952 	/* This interrupt is VMC page fault.*/
953 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC,
954 			      VMC_1_0__SRCID__VM_FAULT,
955 			      &adev->gmc.vm_fault);
956 
957 	if (r)
958 		return r;
959 
960 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2,
961 			      UTCL2_1_0__SRCID__FAULT,
962 			      &adev->gmc.vm_fault);
963 	if (r)
964 		return r;
965 
966 	if (!amdgpu_sriov_vf(adev)) {
967 		/* interrupt sent to DF. */
968 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
969 				      &adev->gmc.ecc_irq);
970 		if (r)
971 			return r;
972 	}
973 
974 	/*
975 	 * Set the internal MC address mask This is the max address of the GPU's
976 	 * internal address space.
977 	 */
978 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
979 
980 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
981 	if (r) {
982 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
983 		return r;
984 	}
985 
986 	adev->need_swiotlb = drm_need_swiotlb(44);
987 
988 	r = gmc_v10_0_mc_init(adev);
989 	if (r)
990 		return r;
991 
992 	amdgpu_gmc_get_vbios_allocations(adev);
993 
994 	/* Memory manager */
995 	r = amdgpu_bo_init(adev);
996 	if (r)
997 		return r;
998 
999 	r = gmc_v10_0_gart_init(adev);
1000 	if (r)
1001 		return r;
1002 
1003 	/*
1004 	 * number of VMs
1005 	 * VMID 0 is reserved for System
1006 	 * amdgpu graphics/compute will use VMIDs 1-7
1007 	 * amdkfd will use VMIDs 8-15
1008 	 */
1009 	adev->vm_manager.first_kfd_vmid = 8;
1010 
1011 	amdgpu_vm_manager_init(adev);
1012 
1013 	return 0;
1014 }
1015 
1016 /**
1017  * gmc_v10_0_gart_fini - vm fini callback
1018  *
1019  * @adev: amdgpu_device pointer
1020  *
1021  * Tears down the driver GART/VM setup (CIK).
1022  */
1023 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev)
1024 {
1025 	amdgpu_gart_table_vram_free(adev);
1026 }
1027 
1028 static int gmc_v10_0_sw_fini(void *handle)
1029 {
1030 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1031 
1032 	amdgpu_vm_manager_fini(adev);
1033 	gmc_v10_0_gart_fini(adev);
1034 	amdgpu_gem_force_release(adev);
1035 	amdgpu_bo_fini(adev);
1036 
1037 	return 0;
1038 }
1039 
1040 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev)
1041 {
1042 }
1043 
1044 /**
1045  * gmc_v10_0_gart_enable - gart enable
1046  *
1047  * @adev: amdgpu_device pointer
1048  */
1049 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev)
1050 {
1051 	int r;
1052 	bool value;
1053 
1054 	if (adev->gart.bo == NULL) {
1055 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1056 		return -EINVAL;
1057 	}
1058 
1059 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
1060 	r = adev->gfxhub.funcs->gart_enable(adev);
1061 	if (r)
1062 		return r;
1063 
1064 	r = adev->mmhub.funcs->gart_enable(adev);
1065 	if (r)
1066 		return r;
1067 
1068 	adev->hdp.funcs->init_registers(adev);
1069 
1070 	/* Flush HDP after it is initialized */
1071 	adev->hdp.funcs->flush_hdp(adev, NULL);
1072 
1073 	value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
1074 		false : true;
1075 
1076 	adev->gfxhub.funcs->set_fault_enable_default(adev, value);
1077 	adev->mmhub.funcs->set_fault_enable_default(adev, value);
1078 	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0);
1079 	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB_0, 0);
1080 
1081 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1082 		 (unsigned)(adev->gmc.gart_size >> 20),
1083 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1084 
1085 	return 0;
1086 }
1087 
1088 static int gmc_v10_0_hw_init(void *handle)
1089 {
1090 	int r;
1091 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1092 
1093 	/* The sequence of these two function calls matters.*/
1094 	gmc_v10_0_init_golden_registers(adev);
1095 
1096 	/*
1097 	 * harvestable groups in gc_utcl2 need to be programmed before any GFX block
1098 	 * register setup within GMC, or else system hang when harvesting SA.
1099 	 */
1100 	if (adev->gfxhub.funcs && adev->gfxhub.funcs->utcl2_harvest)
1101 		adev->gfxhub.funcs->utcl2_harvest(adev);
1102 
1103 	r = gmc_v10_0_gart_enable(adev);
1104 	if (r)
1105 		return r;
1106 
1107 	if (amdgpu_emu_mode == 1) {
1108 		r = amdgpu_gmc_vram_checking(adev);
1109 		if (r)
1110 			return r;
1111 	}
1112 
1113 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
1114 		adev->umc.funcs->init_registers(adev);
1115 
1116 	return 0;
1117 }
1118 
1119 /**
1120  * gmc_v10_0_gart_disable - gart disable
1121  *
1122  * @adev: amdgpu_device pointer
1123  *
1124  * This disables all VM page table.
1125  */
1126 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev)
1127 {
1128 	adev->gfxhub.funcs->gart_disable(adev);
1129 	adev->mmhub.funcs->gart_disable(adev);
1130 }
1131 
1132 static int gmc_v10_0_hw_fini(void *handle)
1133 {
1134 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1135 
1136 	gmc_v10_0_gart_disable(adev);
1137 
1138 	if (amdgpu_sriov_vf(adev)) {
1139 		/* full access mode, so don't touch any GMC register */
1140 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
1141 		return 0;
1142 	}
1143 
1144 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
1145 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
1146 
1147 	return 0;
1148 }
1149 
1150 static int gmc_v10_0_suspend(void *handle)
1151 {
1152 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1153 
1154 	gmc_v10_0_hw_fini(adev);
1155 
1156 	return 0;
1157 }
1158 
1159 static int gmc_v10_0_resume(void *handle)
1160 {
1161 	int r;
1162 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1163 
1164 	r = gmc_v10_0_hw_init(adev);
1165 	if (r)
1166 		return r;
1167 
1168 	amdgpu_vmid_reset_all(adev);
1169 
1170 	return 0;
1171 }
1172 
1173 static bool gmc_v10_0_is_idle(void *handle)
1174 {
1175 	/* MC is always ready in GMC v10.*/
1176 	return true;
1177 }
1178 
1179 static int gmc_v10_0_wait_for_idle(void *handle)
1180 {
1181 	/* There is no need to wait for MC idle in GMC v10.*/
1182 	return 0;
1183 }
1184 
1185 static int gmc_v10_0_soft_reset(void *handle)
1186 {
1187 	return 0;
1188 }
1189 
1190 static int gmc_v10_0_set_clockgating_state(void *handle,
1191 					   enum amd_clockgating_state state)
1192 {
1193 	int r;
1194 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1195 
1196 	/*
1197 	 * The issue mmhub can't disconnect from DF with MMHUB clock gating being disabled
1198 	 * is a new problem observed at DF 3.0.3, however with the same suspend sequence not
1199 	 * seen any issue on the DF 3.0.2 series platform.
1200 	 */
1201 	if (adev->in_s0ix && adev->ip_versions[DF_HWIP][0] > IP_VERSION(3, 0, 2)) {
1202 		dev_dbg(adev->dev, "keep mmhub clock gating being enabled for s0ix\n");
1203 		return 0;
1204 	}
1205 
1206 	r = adev->mmhub.funcs->set_clockgating(adev, state);
1207 	if (r)
1208 		return r;
1209 
1210 	if (adev->ip_versions[ATHUB_HWIP][0] >= IP_VERSION(2, 1, 0))
1211 		return athub_v2_1_set_clockgating(adev, state);
1212 	else
1213 		return athub_v2_0_set_clockgating(adev, state);
1214 }
1215 
1216 static void gmc_v10_0_get_clockgating_state(void *handle, u64 *flags)
1217 {
1218 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1219 
1220 	if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(10, 1, 3) ||
1221 	    adev->ip_versions[GC_HWIP][0] == IP_VERSION(10, 1, 4))
1222 		return;
1223 
1224 	adev->mmhub.funcs->get_clockgating(adev, flags);
1225 
1226 	if (adev->ip_versions[ATHUB_HWIP][0] >= IP_VERSION(2, 1, 0))
1227 		athub_v2_1_get_clockgating(adev, flags);
1228 	else
1229 		athub_v2_0_get_clockgating(adev, flags);
1230 }
1231 
1232 static int gmc_v10_0_set_powergating_state(void *handle,
1233 					   enum amd_powergating_state state)
1234 {
1235 	return 0;
1236 }
1237 
1238 const struct amd_ip_funcs gmc_v10_0_ip_funcs = {
1239 	.name = "gmc_v10_0",
1240 	.early_init = gmc_v10_0_early_init,
1241 	.late_init = gmc_v10_0_late_init,
1242 	.sw_init = gmc_v10_0_sw_init,
1243 	.sw_fini = gmc_v10_0_sw_fini,
1244 	.hw_init = gmc_v10_0_hw_init,
1245 	.hw_fini = gmc_v10_0_hw_fini,
1246 	.suspend = gmc_v10_0_suspend,
1247 	.resume = gmc_v10_0_resume,
1248 	.is_idle = gmc_v10_0_is_idle,
1249 	.wait_for_idle = gmc_v10_0_wait_for_idle,
1250 	.soft_reset = gmc_v10_0_soft_reset,
1251 	.set_clockgating_state = gmc_v10_0_set_clockgating_state,
1252 	.set_powergating_state = gmc_v10_0_set_powergating_state,
1253 	.get_clockgating_state = gmc_v10_0_get_clockgating_state,
1254 };
1255 
1256 const struct amdgpu_ip_block_version gmc_v10_0_ip_block =
1257 {
1258 	.type = AMD_IP_BLOCK_TYPE_GMC,
1259 	.major = 10,
1260 	.minor = 0,
1261 	.rev = 0,
1262 	.funcs = &gmc_v10_0_ip_funcs,
1263 };
1264