xref: /linux/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c (revision 25396684b57f7d16306ca149c545db60b2d08dda)
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 static int gmc_v10_0_ecc_interrupt_state(struct amdgpu_device *adev,
57 					 struct amdgpu_irq_src *src,
58 					 unsigned int type,
59 					 enum amdgpu_interrupt_state state)
60 {
61 	return 0;
62 }
63 
64 static int
65 gmc_v10_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
66 				   struct amdgpu_irq_src *src, unsigned int type,
67 				   enum amdgpu_interrupt_state state)
68 {
69 	switch (state) {
70 	case AMDGPU_IRQ_STATE_DISABLE:
71 		/* MM HUB */
72 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), false);
73 		/* GFX HUB */
74 		/* This works because this interrupt is only
75 		 * enabled at init/resume and disabled in
76 		 * fini/suspend, so the overall state doesn't
77 		 * change over the course of suspend/resume.
78 		 */
79 		if (!adev->in_s0ix)
80 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), false);
81 		break;
82 	case AMDGPU_IRQ_STATE_ENABLE:
83 		/* MM HUB */
84 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), true);
85 		/* GFX HUB */
86 		/* This works because this interrupt is only
87 		 * enabled at init/resume and disabled in
88 		 * fini/suspend, so the overall state doesn't
89 		 * change over the course of suspend/resume.
90 		 */
91 		if (!adev->in_s0ix)
92 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), true);
93 		break;
94 	default:
95 		break;
96 	}
97 
98 	return 0;
99 }
100 
101 static int gmc_v10_0_process_interrupt(struct amdgpu_device *adev,
102 				       struct amdgpu_irq_src *source,
103 				       struct amdgpu_iv_entry *entry)
104 {
105 	uint32_t vmhub_index = entry->client_id == SOC15_IH_CLIENTID_VMC ?
106 			       AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0);
107 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub_index];
108 	bool retry_fault = !!(entry->src_data[1] & 0x80);
109 	bool write_fault = !!(entry->src_data[1] & 0x20);
110 	struct amdgpu_task_info task_info;
111 	uint32_t status = 0;
112 	u64 addr;
113 
114 	addr = (u64)entry->src_data[0] << 12;
115 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
116 
117 	if (retry_fault) {
118 		/* Returning 1 here also prevents sending the IV to the KFD */
119 
120 		/* Process it onyl if it's the first fault for this address */
121 		if (entry->ih != &adev->irq.ih_soft &&
122 		    amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid,
123 					     entry->timestamp))
124 			return 1;
125 
126 		/* Delegate it to a different ring if the hardware hasn't
127 		 * already done it.
128 		 */
129 		if (entry->ih == &adev->irq.ih) {
130 			amdgpu_irq_delegate(adev, entry, 8);
131 			return 1;
132 		}
133 
134 		/* Try to handle the recoverable page faults by filling page
135 		 * tables
136 		 */
137 		if (amdgpu_vm_handle_fault(adev, entry->pasid, 0, 0, addr, write_fault))
138 			return 1;
139 	}
140 
141 	if (!amdgpu_sriov_vf(adev)) {
142 		/*
143 		 * Issue a dummy read to wait for the status register to
144 		 * be updated to avoid reading an incorrect value due to
145 		 * the new fast GRBM interface.
146 		 */
147 		if ((entry->vmid_src == AMDGPU_GFXHUB(0)) &&
148 		    (amdgpu_ip_version(adev, GC_HWIP, 0) <
149 		     IP_VERSION(10, 3, 0)))
150 			RREG32(hub->vm_l2_pro_fault_status);
151 
152 		status = RREG32(hub->vm_l2_pro_fault_status);
153 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
154 	}
155 
156 	if (!printk_ratelimit())
157 		return 0;
158 
159 	memset(&task_info, 0, sizeof(struct amdgpu_task_info));
160 	amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
161 
162 	dev_err(adev->dev,
163 		"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, for process %s pid %d thread %s pid %d)\n",
164 		entry->vmid_src ? "mmhub" : "gfxhub",
165 		entry->src_id, entry->ring_id, entry->vmid,
166 		entry->pasid, task_info.process_name, task_info.tgid,
167 		task_info.task_name, task_info.pid);
168 	dev_err(adev->dev, "  in page starting at address 0x%016llx from client 0x%x (%s)\n",
169 		addr, entry->client_id,
170 		soc15_ih_clientid_name[entry->client_id]);
171 
172 	if (!amdgpu_sriov_vf(adev))
173 		hub->vmhub_funcs->print_l2_protection_fault_status(adev,
174 								   status);
175 
176 	return 0;
177 }
178 
179 static const struct amdgpu_irq_src_funcs gmc_v10_0_irq_funcs = {
180 	.set = gmc_v10_0_vm_fault_interrupt_state,
181 	.process = gmc_v10_0_process_interrupt,
182 };
183 
184 static const struct amdgpu_irq_src_funcs gmc_v10_0_ecc_funcs = {
185 	.set = gmc_v10_0_ecc_interrupt_state,
186 	.process = amdgpu_umc_process_ecc_irq,
187 };
188 
189 static void gmc_v10_0_set_irq_funcs(struct amdgpu_device *adev)
190 {
191 	adev->gmc.vm_fault.num_types = 1;
192 	adev->gmc.vm_fault.funcs = &gmc_v10_0_irq_funcs;
193 
194 	if (!amdgpu_sriov_vf(adev)) {
195 		adev->gmc.ecc_irq.num_types = 1;
196 		adev->gmc.ecc_irq.funcs = &gmc_v10_0_ecc_funcs;
197 	}
198 }
199 
200 /**
201  * gmc_v10_0_use_invalidate_semaphore - judge whether to use semaphore
202  *
203  * @adev: amdgpu_device pointer
204  * @vmhub: vmhub type
205  *
206  */
207 static bool gmc_v10_0_use_invalidate_semaphore(struct amdgpu_device *adev,
208 				       uint32_t vmhub)
209 {
210 	return ((vmhub == AMDGPU_MMHUB0(0)) &&
211 		(!amdgpu_sriov_vf(adev)));
212 }
213 
214 static bool gmc_v10_0_get_atc_vmid_pasid_mapping_info(
215 					struct amdgpu_device *adev,
216 					uint8_t vmid, uint16_t *p_pasid)
217 {
218 	uint32_t value;
219 
220 	value = RREG32(SOC15_REG_OFFSET(ATHUB, 0, mmATC_VMID0_PASID_MAPPING)
221 		     + vmid);
222 	*p_pasid = value & ATC_VMID0_PASID_MAPPING__PASID_MASK;
223 
224 	return !!(value & ATC_VMID0_PASID_MAPPING__VALID_MASK);
225 }
226 
227 /*
228  * GART
229  * VMID 0 is the physical GPU addresses as used by the kernel.
230  * VMIDs 1-15 are used for userspace clients and are handled
231  * by the amdgpu vm/hsa code.
232  */
233 
234 static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
235 				   unsigned int vmhub, uint32_t flush_type)
236 {
237 	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub);
238 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
239 	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
240 	u32 tmp;
241 	/* Use register 17 for GART */
242 	const unsigned int eng = 17;
243 	unsigned int i;
244 	unsigned char hub_ip = 0;
245 
246 	hub_ip = (vmhub == AMDGPU_GFXHUB(0)) ?
247 		   GC_HWIP : MMHUB_HWIP;
248 
249 	spin_lock(&adev->gmc.invalidate_lock);
250 	/*
251 	 * It may lose gpuvm invalidate acknowldege state across power-gating
252 	 * off cycle, add semaphore acquire before invalidation and semaphore
253 	 * release after invalidation to avoid entering power gated state
254 	 * to WA the Issue
255 	 */
256 
257 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
258 	if (use_semaphore) {
259 		for (i = 0; i < adev->usec_timeout; i++) {
260 			/* a read return value of 1 means semaphore acuqire */
261 			tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
262 					 hub->eng_distance * eng, hub_ip);
263 
264 			if (tmp & 0x1)
265 				break;
266 			udelay(1);
267 		}
268 
269 		if (i >= adev->usec_timeout)
270 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
271 	}
272 
273 	WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req +
274 			  hub->eng_distance * eng,
275 			  inv_req, hub_ip);
276 
277 	/*
278 	 * Issue a dummy read to wait for the ACK register to be cleared
279 	 * to avoid a false ACK due to the new fast GRBM interface.
280 	 */
281 	if ((vmhub == AMDGPU_GFXHUB(0)) &&
282 	    (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(10, 3, 0)))
283 		RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req +
284 				  hub->eng_distance * eng, hub_ip);
285 
286 	/* Wait for ACK with a delay.*/
287 	for (i = 0; i < adev->usec_timeout; i++) {
288 		tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack +
289 				  hub->eng_distance * eng, hub_ip);
290 
291 		tmp &= 1 << vmid;
292 		if (tmp)
293 			break;
294 
295 		udelay(1);
296 	}
297 
298 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
299 	if (use_semaphore)
300 		/*
301 		 * add semaphore release after invalidation,
302 		 * write with 0 means semaphore release
303 		 */
304 		WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
305 				  hub->eng_distance * eng, 0, hub_ip);
306 
307 	spin_unlock(&adev->gmc.invalidate_lock);
308 
309 	if (i < adev->usec_timeout)
310 		return;
311 
312 	DRM_ERROR("Timeout waiting for VM flush hub: %d!\n", vmhub);
313 }
314 
315 /**
316  * gmc_v10_0_flush_gpu_tlb - gart tlb flush callback
317  *
318  * @adev: amdgpu_device pointer
319  * @vmid: vm instance to flush
320  * @vmhub: vmhub type
321  * @flush_type: the flush type
322  *
323  * Flush the TLB for the requested page table.
324  */
325 static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
326 					uint32_t vmhub, uint32_t flush_type)
327 {
328 	struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
329 	struct dma_fence *fence;
330 	struct amdgpu_job *job;
331 
332 	int r;
333 
334 	/* flush hdp cache */
335 	adev->hdp.funcs->flush_hdp(adev, NULL);
336 
337 	/* For SRIOV run time, driver shouldn't access the register through MMIO
338 	 * Directly use kiq to do the vm invalidation instead
339 	 */
340 	if (adev->gfx.kiq[0].ring.sched.ready && !adev->enable_mes &&
341 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev)) &&
342 	    down_read_trylock(&adev->reset_domain->sem)) {
343 		struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
344 		const unsigned int eng = 17;
345 		u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
346 		u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
347 		u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
348 
349 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
350 				1 << vmid);
351 
352 		up_read(&adev->reset_domain->sem);
353 		return;
354 	}
355 
356 	mutex_lock(&adev->mman.gtt_window_lock);
357 
358 	if (vmhub == AMDGPU_MMHUB0(0)) {
359 		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_MMHUB0(0), 0);
360 		mutex_unlock(&adev->mman.gtt_window_lock);
361 		return;
362 	}
363 
364 	BUG_ON(vmhub != AMDGPU_GFXHUB(0));
365 
366 	if (!adev->mman.buffer_funcs_enabled ||
367 	    !adev->ib_pool_ready ||
368 	    amdgpu_in_reset(adev) ||
369 	    ring->sched.ready == false) {
370 		gmc_v10_0_flush_vm_hub(adev, vmid, AMDGPU_GFXHUB(0), 0);
371 		mutex_unlock(&adev->mman.gtt_window_lock);
372 		return;
373 	}
374 
375 	/* The SDMA on Navi has a bug which can theoretically result in memory
376 	 * corruption if an invalidation happens at the same time as an VA
377 	 * translation. Avoid this by doing the invalidation from the SDMA
378 	 * itself.
379 	 */
380 	r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.high_pr,
381 				     AMDGPU_FENCE_OWNER_UNDEFINED,
382 				     16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
383 				     &job);
384 	if (r)
385 		goto error_alloc;
386 
387 	job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
388 	job->vm_needs_flush = true;
389 	job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
390 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
391 	fence = amdgpu_job_submit(job);
392 
393 	mutex_unlock(&adev->mman.gtt_window_lock);
394 
395 	dma_fence_wait(fence, false);
396 	dma_fence_put(fence);
397 
398 	return;
399 
400 error_alloc:
401 	mutex_unlock(&adev->mman.gtt_window_lock);
402 	DRM_ERROR("Error flushing GPU TLB using the SDMA (%d)!\n", r);
403 }
404 
405 /**
406  * gmc_v10_0_flush_gpu_tlb_pasid - tlb flush via pasid
407  *
408  * @adev: amdgpu_device pointer
409  * @pasid: pasid to be flush
410  * @flush_type: the flush type
411  * @all_hub: Used with PACKET3_INVALIDATE_TLBS_ALL_HUB()
412  * @inst: is used to select which instance of KIQ to use for the invalidation
413  *
414  * Flush the TLB for the requested pasid.
415  */
416 static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
417 					uint16_t pasid, uint32_t flush_type,
418 					bool all_hub, uint32_t inst)
419 {
420 	int vmid, i;
421 	signed long r;
422 	uint32_t seq;
423 	uint16_t queried_pasid;
424 	bool ret;
425 	u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT : adev->usec_timeout;
426 	struct amdgpu_ring *ring = &adev->gfx.kiq[0].ring;
427 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
428 
429 	if (amdgpu_emu_mode == 0 && ring->sched.ready) {
430 		spin_lock(&adev->gfx.kiq[0].ring_lock);
431 		/* 2 dwords flush + 8 dwords fence */
432 		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8);
433 		kiq->pmf->kiq_invalidate_tlbs(ring,
434 					pasid, flush_type, all_hub);
435 		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
436 		if (r) {
437 			amdgpu_ring_undo(ring);
438 			spin_unlock(&adev->gfx.kiq[0].ring_lock);
439 			return -ETIME;
440 		}
441 
442 		amdgpu_ring_commit(ring);
443 		spin_unlock(&adev->gfx.kiq[0].ring_lock);
444 		r = amdgpu_fence_wait_polling(ring, seq, usec_timeout);
445 		if (r < 1) {
446 			dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
447 			return -ETIME;
448 		}
449 
450 		return 0;
451 	}
452 
453 	for (vmid = 1; vmid < AMDGPU_NUM_VMID; vmid++) {
454 
455 		ret = gmc_v10_0_get_atc_vmid_pasid_mapping_info(adev, vmid,
456 				&queried_pasid);
457 		if (ret	&& queried_pasid == pasid) {
458 			if (all_hub) {
459 				for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS)
460 					gmc_v10_0_flush_gpu_tlb(adev, vmid,
461 							i, flush_type);
462 			} else {
463 				gmc_v10_0_flush_gpu_tlb(adev, vmid,
464 						AMDGPU_GFXHUB(0), flush_type);
465 			}
466 			if (!adev->enable_mes)
467 				break;
468 		}
469 	}
470 
471 	return 0;
472 }
473 
474 static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
475 					     unsigned int vmid, uint64_t pd_addr)
476 {
477 	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(ring->adev, ring->vm_hub);
478 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
479 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
480 	unsigned int eng = ring->vm_inv_eng;
481 
482 	/*
483 	 * It may lose gpuvm invalidate acknowldege state across power-gating
484 	 * off cycle, add semaphore acquire before invalidation and semaphore
485 	 * release after invalidation to avoid entering power gated state
486 	 * to WA the Issue
487 	 */
488 
489 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
490 	if (use_semaphore)
491 		/* a read return value of 1 means semaphore acuqire */
492 		amdgpu_ring_emit_reg_wait(ring,
493 					  hub->vm_inv_eng0_sem +
494 					  hub->eng_distance * eng, 0x1, 0x1);
495 
496 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
497 			      (hub->ctx_addr_distance * vmid),
498 			      lower_32_bits(pd_addr));
499 
500 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
501 			      (hub->ctx_addr_distance * vmid),
502 			      upper_32_bits(pd_addr));
503 
504 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
505 					    hub->eng_distance * eng,
506 					    hub->vm_inv_eng0_ack +
507 					    hub->eng_distance * eng,
508 					    req, 1 << vmid);
509 
510 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
511 	if (use_semaphore)
512 		/*
513 		 * add semaphore release after invalidation,
514 		 * write with 0 means semaphore release
515 		 */
516 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
517 				      hub->eng_distance * eng, 0);
518 
519 	return pd_addr;
520 }
521 
522 static void gmc_v10_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid,
523 					 unsigned int pasid)
524 {
525 	struct amdgpu_device *adev = ring->adev;
526 	uint32_t reg;
527 
528 	/* MES fw manages IH_VMID_x_LUT updating */
529 	if (ring->is_mes_queue)
530 		return;
531 
532 	if (ring->vm_hub == AMDGPU_GFXHUB(0))
533 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT) + vmid;
534 	else
535 		reg = SOC15_REG_OFFSET(OSSSYS, 0, mmIH_VMID_0_LUT_MM) + vmid;
536 
537 	amdgpu_ring_emit_wreg(ring, reg, pasid);
538 }
539 
540 /*
541  * PTE format on NAVI 10:
542  * 63:59 reserved
543  * 58 reserved and for sienna_cichlid is used for MALL noalloc
544  * 57 reserved
545  * 56 F
546  * 55 L
547  * 54 reserved
548  * 53:52 SW
549  * 51 T
550  * 50:48 mtype
551  * 47:12 4k physical page base address
552  * 11:7 fragment
553  * 6 write
554  * 5 read
555  * 4 exe
556  * 3 Z
557  * 2 snooped
558  * 1 system
559  * 0 valid
560  *
561  * PDE format on NAVI 10:
562  * 63:59 block fragment size
563  * 58:55 reserved
564  * 54 P
565  * 53:48 reserved
566  * 47:6 physical base address of PD or PTE
567  * 5:3 reserved
568  * 2 C
569  * 1 system
570  * 0 valid
571  */
572 
573 static uint64_t gmc_v10_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
574 {
575 	switch (flags) {
576 	case AMDGPU_VM_MTYPE_DEFAULT:
577 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
578 	case AMDGPU_VM_MTYPE_NC:
579 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
580 	case AMDGPU_VM_MTYPE_WC:
581 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
582 	case AMDGPU_VM_MTYPE_CC:
583 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
584 	case AMDGPU_VM_MTYPE_UC:
585 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
586 	default:
587 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
588 	}
589 }
590 
591 static void gmc_v10_0_get_vm_pde(struct amdgpu_device *adev, int level,
592 				 uint64_t *addr, uint64_t *flags)
593 {
594 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
595 		*addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
596 	BUG_ON(*addr & 0xFFFF00000000003FULL);
597 
598 	if (!adev->gmc.translate_further)
599 		return;
600 
601 	if (level == AMDGPU_VM_PDB1) {
602 		/* Set the block fragment size */
603 		if (!(*flags & AMDGPU_PDE_PTE))
604 			*flags |= AMDGPU_PDE_BFS(0x9);
605 
606 	} else if (level == AMDGPU_VM_PDB0) {
607 		if (*flags & AMDGPU_PDE_PTE)
608 			*flags &= ~AMDGPU_PDE_PTE;
609 		else
610 			*flags |= AMDGPU_PTE_TF;
611 	}
612 }
613 
614 static void gmc_v10_0_get_vm_pte(struct amdgpu_device *adev,
615 				 struct amdgpu_bo_va_mapping *mapping,
616 				 uint64_t *flags)
617 {
618 	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
619 
620 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
621 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
622 
623 	*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
624 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
625 
626 	*flags &= ~AMDGPU_PTE_NOALLOC;
627 	*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
628 
629 	if (mapping->flags & AMDGPU_PTE_PRT) {
630 		*flags |= AMDGPU_PTE_PRT;
631 		*flags |= AMDGPU_PTE_SNOOPED;
632 		*flags |= AMDGPU_PTE_LOG;
633 		*flags |= AMDGPU_PTE_SYSTEM;
634 		*flags &= ~AMDGPU_PTE_VALID;
635 	}
636 
637 	if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
638 			       AMDGPU_GEM_CREATE_EXT_COHERENT |
639 			       AMDGPU_GEM_CREATE_UNCACHED))
640 		*flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
641 			 AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
642 }
643 
644 static unsigned int gmc_v10_0_get_vbios_fb_size(struct amdgpu_device *adev)
645 {
646 	u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
647 	unsigned int size;
648 
649 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
650 		size = AMDGPU_VBIOS_VGA_ALLOCATION;
651 	} else {
652 		u32 viewport;
653 		u32 pitch;
654 
655 		viewport = RREG32_SOC15(DCE, 0, mmHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
656 		pitch = RREG32_SOC15(DCE, 0, mmHUBPREQ0_DCSURF_SURFACE_PITCH);
657 		size = (REG_GET_FIELD(viewport,
658 					HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
659 				REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) *
660 				4);
661 	}
662 
663 	return size;
664 }
665 
666 static const struct amdgpu_gmc_funcs gmc_v10_0_gmc_funcs = {
667 	.flush_gpu_tlb = gmc_v10_0_flush_gpu_tlb,
668 	.flush_gpu_tlb_pasid = gmc_v10_0_flush_gpu_tlb_pasid,
669 	.emit_flush_gpu_tlb = gmc_v10_0_emit_flush_gpu_tlb,
670 	.emit_pasid_mapping = gmc_v10_0_emit_pasid_mapping,
671 	.map_mtype = gmc_v10_0_map_mtype,
672 	.get_vm_pde = gmc_v10_0_get_vm_pde,
673 	.get_vm_pte = gmc_v10_0_get_vm_pte,
674 	.get_vbios_fb_size = gmc_v10_0_get_vbios_fb_size,
675 };
676 
677 static void gmc_v10_0_set_gmc_funcs(struct amdgpu_device *adev)
678 {
679 	if (adev->gmc.gmc_funcs == NULL)
680 		adev->gmc.gmc_funcs = &gmc_v10_0_gmc_funcs;
681 }
682 
683 static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev)
684 {
685 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
686 	case IP_VERSION(8, 7, 0):
687 		adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM;
688 		adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM;
689 		adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM;
690 		adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA;
691 		adev->umc.retire_unit = 1;
692 		adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0];
693 		adev->umc.ras = &umc_v8_7_ras;
694 		break;
695 	default:
696 		break;
697 	}
698 }
699 
700 static void gmc_v10_0_set_mmhub_funcs(struct amdgpu_device *adev)
701 {
702 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
703 	case IP_VERSION(2, 3, 0):
704 	case IP_VERSION(2, 4, 0):
705 	case IP_VERSION(2, 4, 1):
706 		adev->mmhub.funcs = &mmhub_v2_3_funcs;
707 		break;
708 	default:
709 		adev->mmhub.funcs = &mmhub_v2_0_funcs;
710 		break;
711 	}
712 }
713 
714 static void gmc_v10_0_set_gfxhub_funcs(struct amdgpu_device *adev)
715 {
716 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
717 	case IP_VERSION(10, 3, 0):
718 	case IP_VERSION(10, 3, 2):
719 	case IP_VERSION(10, 3, 1):
720 	case IP_VERSION(10, 3, 4):
721 	case IP_VERSION(10, 3, 5):
722 	case IP_VERSION(10, 3, 6):
723 	case IP_VERSION(10, 3, 3):
724 	case IP_VERSION(10, 3, 7):
725 		adev->gfxhub.funcs = &gfxhub_v2_1_funcs;
726 		break;
727 	default:
728 		adev->gfxhub.funcs = &gfxhub_v2_0_funcs;
729 		break;
730 	}
731 }
732 
733 
734 static int gmc_v10_0_early_init(void *handle)
735 {
736 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
737 
738 	gmc_v10_0_set_mmhub_funcs(adev);
739 	gmc_v10_0_set_gfxhub_funcs(adev);
740 	gmc_v10_0_set_gmc_funcs(adev);
741 	gmc_v10_0_set_irq_funcs(adev);
742 	gmc_v10_0_set_umc_funcs(adev);
743 
744 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
745 	adev->gmc.shared_aperture_end =
746 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
747 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
748 	adev->gmc.private_aperture_end =
749 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
750 	adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF;
751 
752 	return 0;
753 }
754 
755 static int gmc_v10_0_late_init(void *handle)
756 {
757 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
758 	int r;
759 
760 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
761 	if (r)
762 		return r;
763 
764 	r = amdgpu_gmc_ras_late_init(adev);
765 	if (r)
766 		return r;
767 
768 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
769 }
770 
771 static void gmc_v10_0_vram_gtt_location(struct amdgpu_device *adev,
772 					struct amdgpu_gmc *mc)
773 {
774 	u64 base = 0;
775 
776 	base = adev->gfxhub.funcs->get_fb_location(adev);
777 
778 	/* add the xgmi offset of the physical node */
779 	base += adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
780 
781 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
782 	amdgpu_gmc_gart_location(adev, mc);
783 	amdgpu_gmc_agp_location(adev, mc);
784 
785 	/* base offset of vram pages */
786 	adev->vm_manager.vram_base_offset = adev->gfxhub.funcs->get_mc_fb_offset(adev);
787 
788 	/* add the xgmi offset of the physical node */
789 	adev->vm_manager.vram_base_offset +=
790 		adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
791 }
792 
793 /**
794  * gmc_v10_0_mc_init - initialize the memory controller driver params
795  *
796  * @adev: amdgpu_device pointer
797  *
798  * Look up the amount of vram, vram width, and decide how to place
799  * vram and gart within the GPU's physical address space.
800  * Returns 0 for success.
801  */
802 static int gmc_v10_0_mc_init(struct amdgpu_device *adev)
803 {
804 	int r;
805 
806 	/* size in MB on si */
807 	adev->gmc.mc_vram_size =
808 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
809 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
810 
811 	if (!(adev->flags & AMD_IS_APU)) {
812 		r = amdgpu_device_resize_fb_bar(adev);
813 		if (r)
814 			return r;
815 	}
816 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
817 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
818 
819 #ifdef CONFIG_X86_64
820 	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) {
821 		adev->gmc.aper_base = adev->gfxhub.funcs->get_mc_fb_offset(adev);
822 		adev->gmc.aper_size = adev->gmc.real_vram_size;
823 	}
824 #endif
825 
826 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
827 
828 	/* set the gart size */
829 	if (amdgpu_gart_size == -1) {
830 		switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
831 		default:
832 			adev->gmc.gart_size = 512ULL << 20;
833 			break;
834 		case IP_VERSION(10, 3, 1):   /* DCE SG support */
835 		case IP_VERSION(10, 3, 3):   /* DCE SG support */
836 		case IP_VERSION(10, 3, 6):   /* DCE SG support */
837 		case IP_VERSION(10, 3, 7):   /* DCE SG support */
838 			adev->gmc.gart_size = 1024ULL << 20;
839 			break;
840 		}
841 	} else {
842 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
843 	}
844 
845 	gmc_v10_0_vram_gtt_location(adev, &adev->gmc);
846 
847 	return 0;
848 }
849 
850 static int gmc_v10_0_gart_init(struct amdgpu_device *adev)
851 {
852 	int r;
853 
854 	if (adev->gart.bo) {
855 		WARN(1, "NAVI10 PCIE GART already initialized\n");
856 		return 0;
857 	}
858 
859 	/* Initialize common gart structure */
860 	r = amdgpu_gart_init(adev);
861 	if (r)
862 		return r;
863 
864 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
865 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
866 				 AMDGPU_PTE_EXECUTABLE;
867 
868 	return amdgpu_gart_table_vram_alloc(adev);
869 }
870 
871 static int gmc_v10_0_sw_init(void *handle)
872 {
873 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
874 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
875 
876 	adev->gfxhub.funcs->init(adev);
877 
878 	adev->mmhub.funcs->init(adev);
879 
880 	spin_lock_init(&adev->gmc.invalidate_lock);
881 
882 	if ((adev->flags & AMD_IS_APU) && amdgpu_emu_mode == 1) {
883 		adev->gmc.vram_type = AMDGPU_VRAM_TYPE_DDR4;
884 		adev->gmc.vram_width = 64;
885 	} else if (amdgpu_emu_mode == 1) {
886 		adev->gmc.vram_type = AMDGPU_VRAM_TYPE_GDDR6;
887 		adev->gmc.vram_width = 1 * 128; /* numchan * chansize */
888 	} else {
889 		r = amdgpu_atomfirmware_get_vram_info(adev,
890 				&vram_width, &vram_type, &vram_vendor);
891 		adev->gmc.vram_width = vram_width;
892 
893 		adev->gmc.vram_type = vram_type;
894 		adev->gmc.vram_vendor = vram_vendor;
895 	}
896 
897 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
898 	case IP_VERSION(10, 3, 0):
899 		adev->gmc.mall_size = 128 * 1024 * 1024;
900 		break;
901 	case IP_VERSION(10, 3, 2):
902 		adev->gmc.mall_size = 96 * 1024 * 1024;
903 		break;
904 	case IP_VERSION(10, 3, 4):
905 		adev->gmc.mall_size = 32 * 1024 * 1024;
906 		break;
907 	case IP_VERSION(10, 3, 5):
908 		adev->gmc.mall_size = 16 * 1024 * 1024;
909 		break;
910 	default:
911 		adev->gmc.mall_size = 0;
912 		break;
913 	}
914 
915 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
916 	case IP_VERSION(10, 1, 10):
917 	case IP_VERSION(10, 1, 1):
918 	case IP_VERSION(10, 1, 2):
919 	case IP_VERSION(10, 1, 3):
920 	case IP_VERSION(10, 1, 4):
921 	case IP_VERSION(10, 3, 0):
922 	case IP_VERSION(10, 3, 2):
923 	case IP_VERSION(10, 3, 1):
924 	case IP_VERSION(10, 3, 4):
925 	case IP_VERSION(10, 3, 5):
926 	case IP_VERSION(10, 3, 6):
927 	case IP_VERSION(10, 3, 3):
928 	case IP_VERSION(10, 3, 7):
929 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
930 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
931 		/*
932 		 * To fulfill 4-level page support,
933 		 * vm size is 256TB (48bit), maximum size of Navi10/Navi14/Navi12,
934 		 * block size 512 (9bit)
935 		 */
936 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
937 		break;
938 	default:
939 		break;
940 	}
941 
942 	/* This interrupt is VMC page fault.*/
943 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_VMC,
944 			      VMC_1_0__SRCID__VM_FAULT,
945 			      &adev->gmc.vm_fault);
946 
947 	if (r)
948 		return r;
949 
950 	r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_UTCL2,
951 			      UTCL2_1_0__SRCID__FAULT,
952 			      &adev->gmc.vm_fault);
953 	if (r)
954 		return r;
955 
956 	if (!amdgpu_sriov_vf(adev)) {
957 		/* interrupt sent to DF. */
958 		r = amdgpu_irq_add_id(adev, SOC15_IH_CLIENTID_DF, 0,
959 				      &adev->gmc.ecc_irq);
960 		if (r)
961 			return r;
962 	}
963 
964 	/*
965 	 * Set the internal MC address mask This is the max address of the GPU's
966 	 * internal address space.
967 	 */
968 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
969 
970 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
971 	if (r) {
972 		dev_warn(adev->dev, "amdgpu: No suitable DMA available.\n");
973 		return r;
974 	}
975 
976 	adev->need_swiotlb = drm_need_swiotlb(44);
977 
978 	r = gmc_v10_0_mc_init(adev);
979 	if (r)
980 		return r;
981 
982 	amdgpu_gmc_get_vbios_allocations(adev);
983 
984 	/* Memory manager */
985 	r = amdgpu_bo_init(adev);
986 	if (r)
987 		return r;
988 
989 	r = gmc_v10_0_gart_init(adev);
990 	if (r)
991 		return r;
992 
993 	/*
994 	 * number of VMs
995 	 * VMID 0 is reserved for System
996 	 * amdgpu graphics/compute will use VMIDs 1-7
997 	 * amdkfd will use VMIDs 8-15
998 	 */
999 	adev->vm_manager.first_kfd_vmid = 8;
1000 
1001 	amdgpu_vm_manager_init(adev);
1002 
1003 	r = amdgpu_gmc_ras_sw_init(adev);
1004 	if (r)
1005 		return r;
1006 
1007 	return 0;
1008 }
1009 
1010 /**
1011  * gmc_v10_0_gart_fini - vm fini callback
1012  *
1013  * @adev: amdgpu_device pointer
1014  *
1015  * Tears down the driver GART/VM setup (CIK).
1016  */
1017 static void gmc_v10_0_gart_fini(struct amdgpu_device *adev)
1018 {
1019 	amdgpu_gart_table_vram_free(adev);
1020 }
1021 
1022 static int gmc_v10_0_sw_fini(void *handle)
1023 {
1024 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1025 
1026 	amdgpu_vm_manager_fini(adev);
1027 	gmc_v10_0_gart_fini(adev);
1028 	amdgpu_gem_force_release(adev);
1029 	amdgpu_bo_fini(adev);
1030 
1031 	return 0;
1032 }
1033 
1034 static void gmc_v10_0_init_golden_registers(struct amdgpu_device *adev)
1035 {
1036 }
1037 
1038 /**
1039  * gmc_v10_0_gart_enable - gart enable
1040  *
1041  * @adev: amdgpu_device pointer
1042  */
1043 static int gmc_v10_0_gart_enable(struct amdgpu_device *adev)
1044 {
1045 	int r;
1046 	bool value;
1047 
1048 	if (adev->gart.bo == NULL) {
1049 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
1050 		return -EINVAL;
1051 	}
1052 
1053 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
1054 
1055 	if (!adev->in_s0ix) {
1056 		r = adev->gfxhub.funcs->gart_enable(adev);
1057 		if (r)
1058 			return r;
1059 	}
1060 
1061 	r = adev->mmhub.funcs->gart_enable(adev);
1062 	if (r)
1063 		return r;
1064 
1065 	adev->hdp.funcs->init_registers(adev);
1066 
1067 	/* Flush HDP after it is initialized */
1068 	adev->hdp.funcs->flush_hdp(adev, NULL);
1069 
1070 	value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
1071 		false : true;
1072 
1073 	if (!adev->in_s0ix)
1074 		adev->gfxhub.funcs->set_fault_enable_default(adev, value);
1075 	adev->mmhub.funcs->set_fault_enable_default(adev, value);
1076 	gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB0(0), 0);
1077 	if (!adev->in_s0ix)
1078 		gmc_v10_0_flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB(0), 0);
1079 
1080 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
1081 		 (unsigned int)(adev->gmc.gart_size >> 20),
1082 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
1083 
1084 	return 0;
1085 }
1086 
1087 static int gmc_v10_0_hw_init(void *handle)
1088 {
1089 	int r;
1090 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1091 
1092 	/* The sequence of these two function calls matters.*/
1093 	gmc_v10_0_init_golden_registers(adev);
1094 
1095 	/*
1096 	 * harvestable groups in gc_utcl2 need to be programmed before any GFX block
1097 	 * register setup within GMC, or else system hang when harvesting SA.
1098 	 */
1099 	if (!adev->in_s0ix && adev->gfxhub.funcs && adev->gfxhub.funcs->utcl2_harvest)
1100 		adev->gfxhub.funcs->utcl2_harvest(adev);
1101 
1102 	r = gmc_v10_0_gart_enable(adev);
1103 	if (r)
1104 		return r;
1105 
1106 	if (amdgpu_emu_mode == 1) {
1107 		r = amdgpu_gmc_vram_checking(adev);
1108 		if (r)
1109 			return r;
1110 	}
1111 
1112 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
1113 		adev->umc.funcs->init_registers(adev);
1114 
1115 	return 0;
1116 }
1117 
1118 /**
1119  * gmc_v10_0_gart_disable - gart disable
1120  *
1121  * @adev: amdgpu_device pointer
1122  *
1123  * This disables all VM page table.
1124  */
1125 static void gmc_v10_0_gart_disable(struct amdgpu_device *adev)
1126 {
1127 	if (!adev->in_s0ix)
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.vm_fault, 0);
1145 
1146 	return 0;
1147 }
1148 
1149 static int gmc_v10_0_suspend(void *handle)
1150 {
1151 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1152 
1153 	gmc_v10_0_hw_fini(adev);
1154 
1155 	return 0;
1156 }
1157 
1158 static int gmc_v10_0_resume(void *handle)
1159 {
1160 	int r;
1161 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1162 
1163 	r = gmc_v10_0_hw_init(adev);
1164 	if (r)
1165 		return r;
1166 
1167 	amdgpu_vmid_reset_all(adev);
1168 
1169 	return 0;
1170 }
1171 
1172 static bool gmc_v10_0_is_idle(void *handle)
1173 {
1174 	/* MC is always ready in GMC v10.*/
1175 	return true;
1176 }
1177 
1178 static int gmc_v10_0_wait_for_idle(void *handle)
1179 {
1180 	/* There is no need to wait for MC idle in GMC v10.*/
1181 	return 0;
1182 }
1183 
1184 static int gmc_v10_0_soft_reset(void *handle)
1185 {
1186 	return 0;
1187 }
1188 
1189 static int gmc_v10_0_set_clockgating_state(void *handle,
1190 					   enum amd_clockgating_state state)
1191 {
1192 	int r;
1193 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1194 
1195 	/*
1196 	 * The issue mmhub can't disconnect from DF with MMHUB clock gating being disabled
1197 	 * is a new problem observed at DF 3.0.3, however with the same suspend sequence not
1198 	 * seen any issue on the DF 3.0.2 series platform.
1199 	 */
1200 	if (adev->in_s0ix &&
1201 	    amdgpu_ip_version(adev, 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 (amdgpu_ip_version(adev, 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 (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(10, 1, 3) ||
1221 	    amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(10, 1, 4))
1222 		return;
1223 
1224 	adev->mmhub.funcs->get_clockgating(adev, flags);
1225 
1226 	if (amdgpu_ip_version(adev, 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 	.type = AMD_IP_BLOCK_TYPE_GMC,
1258 	.major = 10,
1259 	.minor = 0,
1260 	.rev = 0,
1261 	.funcs = &gmc_v10_0_ip_funcs,
1262 };
1263