1 /*
2  * Copyright 2021 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_v11_0.h"
28 #include "umc_v8_7.h"
29 #include "athub/athub_3_0_0_sh_mask.h"
30 #include "athub/athub_3_0_0_offset.h"
31 #include "oss/osssys_6_0_0_offset.h"
32 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
33 #include "navi10_enum.h"
34 #include "soc15.h"
35 #include "soc15d.h"
36 #include "soc15_common.h"
37 #include "nbio_v4_3.h"
38 #include "gfxhub_v3_0.h"
39 #include "mmhub_v3_0.h"
40 #include "mmhub_v3_0_2.h"
41 #include "athub_v3_0.h"
42 
43 
44 static int gmc_v11_0_ecc_interrupt_state(struct amdgpu_device *adev,
45 					 struct amdgpu_irq_src *src,
46 					 unsigned type,
47 					 enum amdgpu_interrupt_state state)
48 {
49 	return 0;
50 }
51 
52 static int
53 gmc_v11_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
54 				   struct amdgpu_irq_src *src, unsigned type,
55 				   enum amdgpu_interrupt_state state)
56 {
57 	switch (state) {
58 	case AMDGPU_IRQ_STATE_DISABLE:
59 		/* MM HUB */
60 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, false);
61 		/* GFX HUB */
62 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, false);
63 		break;
64 	case AMDGPU_IRQ_STATE_ENABLE:
65 		/* MM HUB */
66 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, true);
67 		/* GFX HUB */
68 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, true);
69 		break;
70 	default:
71 		break;
72 	}
73 
74 	return 0;
75 }
76 
77 static int gmc_v11_0_process_interrupt(struct amdgpu_device *adev,
78 				       struct amdgpu_irq_src *source,
79 				       struct amdgpu_iv_entry *entry)
80 {
81 	struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
82 	uint32_t status = 0;
83 	u64 addr;
84 
85 	addr = (u64)entry->src_data[0] << 12;
86 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
87 
88 	if (!amdgpu_sriov_vf(adev)) {
89 		/*
90 		 * Issue a dummy read to wait for the status register to
91 		 * be updated to avoid reading an incorrect value due to
92 		 * the new fast GRBM interface.
93 		 */
94 		if (entry->vmid_src == AMDGPU_GFXHUB_0)
95 			RREG32(hub->vm_l2_pro_fault_status);
96 
97 		status = RREG32(hub->vm_l2_pro_fault_status);
98 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
99 	}
100 
101 	if (printk_ratelimit()) {
102 		struct amdgpu_task_info task_info;
103 
104 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
105 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
106 
107 		dev_err(adev->dev,
108 			"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, "
109 			"for process %s pid %d thread %s pid %d)\n",
110 			entry->vmid_src ? "mmhub" : "gfxhub",
111 			entry->src_id, entry->ring_id, entry->vmid,
112 			entry->pasid, task_info.process_name, task_info.tgid,
113 			task_info.task_name, task_info.pid);
114 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
115 			addr, entry->client_id);
116 		if (!amdgpu_sriov_vf(adev))
117 			hub->vmhub_funcs->print_l2_protection_fault_status(adev, status);
118 	}
119 
120 	return 0;
121 }
122 
123 static const struct amdgpu_irq_src_funcs gmc_v11_0_irq_funcs = {
124 	.set = gmc_v11_0_vm_fault_interrupt_state,
125 	.process = gmc_v11_0_process_interrupt,
126 };
127 
128 static const struct amdgpu_irq_src_funcs gmc_v11_0_ecc_funcs = {
129 	.set = gmc_v11_0_ecc_interrupt_state,
130 	.process = amdgpu_umc_process_ecc_irq,
131 };
132 
133 static void gmc_v11_0_set_irq_funcs(struct amdgpu_device *adev)
134 {
135 	adev->gmc.vm_fault.num_types = 1;
136 	adev->gmc.vm_fault.funcs = &gmc_v11_0_irq_funcs;
137 
138 	if (!amdgpu_sriov_vf(adev)) {
139 		adev->gmc.ecc_irq.num_types = 1;
140 		adev->gmc.ecc_irq.funcs = &gmc_v11_0_ecc_funcs;
141 	}
142 }
143 
144 /**
145  * gmc_v11_0_use_invalidate_semaphore - judge whether to use semaphore
146  *
147  * @adev: amdgpu_device pointer
148  * @vmhub: vmhub type
149  *
150  */
151 static bool gmc_v11_0_use_invalidate_semaphore(struct amdgpu_device *adev,
152 				       uint32_t vmhub)
153 {
154 	return ((vmhub == AMDGPU_MMHUB_0) &&
155 		(!amdgpu_sriov_vf(adev)));
156 }
157 
158 static bool gmc_v11_0_get_vmid_pasid_mapping_info(
159 					struct amdgpu_device *adev,
160 					uint8_t vmid, uint16_t *p_pasid)
161 {
162 	*p_pasid = RREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid) & 0xffff;
163 
164 	return !!(*p_pasid);
165 }
166 
167 /*
168  * GART
169  * VMID 0 is the physical GPU addresses as used by the kernel.
170  * VMIDs 1-15 are used for userspace clients and are handled
171  * by the amdgpu vm/hsa code.
172  */
173 
174 static void gmc_v11_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
175 				   unsigned int vmhub, uint32_t flush_type)
176 {
177 	bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(adev, vmhub);
178 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
179 	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
180 	u32 tmp;
181 	/* Use register 17 for GART */
182 	const unsigned eng = 17;
183 	unsigned int i;
184 
185 	spin_lock(&adev->gmc.invalidate_lock);
186 	/*
187 	 * It may lose gpuvm invalidate acknowldege state across power-gating
188 	 * off cycle, add semaphore acquire before invalidation and semaphore
189 	 * release after invalidation to avoid entering power gated state
190 	 * to WA the Issue
191 	 */
192 
193 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
194 	if (use_semaphore) {
195 		for (i = 0; i < adev->usec_timeout; i++) {
196 			/* a read return value of 1 means semaphore acuqire */
197 			tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_sem +
198 					    hub->eng_distance * eng);
199 			if (tmp & 0x1)
200 				break;
201 			udelay(1);
202 		}
203 
204 		if (i >= adev->usec_timeout)
205 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
206 	}
207 
208 	WREG32_NO_KIQ(hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req);
209 
210 	/* Wait for ACK with a delay.*/
211 	for (i = 0; i < adev->usec_timeout; i++) {
212 		tmp = RREG32_NO_KIQ(hub->vm_inv_eng0_ack +
213 				    hub->eng_distance * eng);
214 		tmp &= 1 << vmid;
215 		if (tmp)
216 			break;
217 
218 		udelay(1);
219 	}
220 
221 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
222 	if (use_semaphore)
223 		/*
224 		 * add semaphore release after invalidation,
225 		 * write with 0 means semaphore release
226 		 */
227 		WREG32_NO_KIQ(hub->vm_inv_eng0_sem +
228 			      hub->eng_distance * eng, 0);
229 
230 	/* Issue additional private vm invalidation to MMHUB */
231 	if ((vmhub != AMDGPU_GFXHUB_0) &&
232 	    (hub->vm_l2_bank_select_reserved_cid2)) {
233 		inv_req = RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
234 		/* bit 25: RSERVED_CACHE_PRIVATE_INVALIDATION */
235 		inv_req |= (1 << 25);
236 		/* Issue private invalidation */
237 		WREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2, inv_req);
238 		/* Read back to ensure invalidation is done*/
239 		RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
240 	}
241 
242 	spin_unlock(&adev->gmc.invalidate_lock);
243 
244 	if (i < adev->usec_timeout)
245 		return;
246 
247 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
248 }
249 
250 /**
251  * gmc_v11_0_flush_gpu_tlb - gart tlb flush callback
252  *
253  * @adev: amdgpu_device pointer
254  * @vmid: vm instance to flush
255  *
256  * Flush the TLB for the requested page table.
257  */
258 static void gmc_v11_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
259 					uint32_t vmhub, uint32_t flush_type)
260 {
261 	if ((vmhub == AMDGPU_GFXHUB_0) && !adev->gfx.is_poweron)
262 		return;
263 
264 	/* flush hdp cache */
265 	adev->hdp.funcs->flush_hdp(adev, NULL);
266 
267 	/* For SRIOV run time, driver shouldn't access the register through MMIO
268 	 * Directly use kiq to do the vm invalidation instead
269 	 */
270 	if (adev->gfx.kiq.ring.sched.ready && !adev->enable_mes &&
271 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) {
272 		struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
273 		const unsigned eng = 17;
274 		u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
275 		u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
276 		u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
277 
278 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
279 				1 << vmid);
280 		return;
281 	}
282 
283 	mutex_lock(&adev->mman.gtt_window_lock);
284 	gmc_v11_0_flush_vm_hub(adev, vmid, vmhub, 0);
285 	mutex_unlock(&adev->mman.gtt_window_lock);
286 	return;
287 }
288 
289 /**
290  * gmc_v11_0_flush_gpu_tlb_pasid - tlb flush via pasid
291  *
292  * @adev: amdgpu_device pointer
293  * @pasid: pasid to be flush
294  *
295  * Flush the TLB for the requested pasid.
296  */
297 static int gmc_v11_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
298 					uint16_t pasid, uint32_t flush_type,
299 					bool all_hub)
300 {
301 	int vmid, i;
302 	signed long r;
303 	uint32_t seq;
304 	uint16_t queried_pasid;
305 	bool ret;
306 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
307 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
308 
309 	if (amdgpu_emu_mode == 0 && ring->sched.ready) {
310 		spin_lock(&adev->gfx.kiq.ring_lock);
311 		/* 2 dwords flush + 8 dwords fence */
312 		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8);
313 		kiq->pmf->kiq_invalidate_tlbs(ring,
314 					pasid, flush_type, all_hub);
315 		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
316 		if (r) {
317 			amdgpu_ring_undo(ring);
318 			spin_unlock(&adev->gfx.kiq.ring_lock);
319 			return -ETIME;
320 		}
321 
322 		amdgpu_ring_commit(ring);
323 		spin_unlock(&adev->gfx.kiq.ring_lock);
324 		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
325 		if (r < 1) {
326 			dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
327 			return -ETIME;
328 		}
329 
330 		return 0;
331 	}
332 
333 	for (vmid = 1; vmid < 16; vmid++) {
334 
335 		ret = gmc_v11_0_get_vmid_pasid_mapping_info(adev, vmid,
336 				&queried_pasid);
337 		if (ret	&& queried_pasid == pasid) {
338 			if (all_hub) {
339 				for (i = 0; i < adev->num_vmhubs; i++)
340 					gmc_v11_0_flush_gpu_tlb(adev, vmid,
341 							i, flush_type);
342 			} else {
343 				gmc_v11_0_flush_gpu_tlb(adev, vmid,
344 						AMDGPU_GFXHUB_0, flush_type);
345 			}
346 			break;
347 		}
348 	}
349 
350 	return 0;
351 }
352 
353 static uint64_t gmc_v11_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
354 					     unsigned vmid, uint64_t pd_addr)
355 {
356 	bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(ring->adev, ring->funcs->vmhub);
357 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->funcs->vmhub];
358 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
359 	unsigned eng = ring->vm_inv_eng;
360 
361 	/*
362 	 * It may lose gpuvm invalidate acknowldege state across power-gating
363 	 * off cycle, add semaphore acquire before invalidation and semaphore
364 	 * release after invalidation to avoid entering power gated state
365 	 * to WA the Issue
366 	 */
367 
368 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
369 	if (use_semaphore)
370 		/* a read return value of 1 means semaphore acuqire */
371 		amdgpu_ring_emit_reg_wait(ring,
372 					  hub->vm_inv_eng0_sem +
373 					  hub->eng_distance * eng, 0x1, 0x1);
374 
375 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
376 			      (hub->ctx_addr_distance * vmid),
377 			      lower_32_bits(pd_addr));
378 
379 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
380 			      (hub->ctx_addr_distance * vmid),
381 			      upper_32_bits(pd_addr));
382 
383 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
384 					    hub->eng_distance * eng,
385 					    hub->vm_inv_eng0_ack +
386 					    hub->eng_distance * eng,
387 					    req, 1 << vmid);
388 
389 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
390 	if (use_semaphore)
391 		/*
392 		 * add semaphore release after invalidation,
393 		 * write with 0 means semaphore release
394 		 */
395 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
396 				      hub->eng_distance * eng, 0);
397 
398 	return pd_addr;
399 }
400 
401 static void gmc_v11_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
402 					 unsigned pasid)
403 {
404 	struct amdgpu_device *adev = ring->adev;
405 	uint32_t reg;
406 
407 	/* MES fw manages IH_VMID_x_LUT updating */
408 	if (ring->is_mes_queue)
409 		return;
410 
411 	if (ring->funcs->vmhub == AMDGPU_GFXHUB_0)
412 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid;
413 	else
414 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT_MM) + vmid;
415 
416 	amdgpu_ring_emit_wreg(ring, reg, pasid);
417 }
418 
419 /*
420  * PTE format:
421  * 63:59 reserved
422  * 58:57 reserved
423  * 56 F
424  * 55 L
425  * 54 reserved
426  * 53:52 SW
427  * 51 T
428  * 50:48 mtype
429  * 47:12 4k physical page base address
430  * 11:7 fragment
431  * 6 write
432  * 5 read
433  * 4 exe
434  * 3 Z
435  * 2 snooped
436  * 1 system
437  * 0 valid
438  *
439  * PDE format:
440  * 63:59 block fragment size
441  * 58:55 reserved
442  * 54 P
443  * 53:48 reserved
444  * 47:6 physical base address of PD or PTE
445  * 5:3 reserved
446  * 2 C
447  * 1 system
448  * 0 valid
449  */
450 
451 static uint64_t gmc_v11_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
452 {
453 	switch (flags) {
454 	case AMDGPU_VM_MTYPE_DEFAULT:
455 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
456 	case AMDGPU_VM_MTYPE_NC:
457 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
458 	case AMDGPU_VM_MTYPE_WC:
459 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
460 	case AMDGPU_VM_MTYPE_CC:
461 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
462 	case AMDGPU_VM_MTYPE_UC:
463 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
464 	default:
465 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
466 	}
467 }
468 
469 static void gmc_v11_0_get_vm_pde(struct amdgpu_device *adev, int level,
470 				 uint64_t *addr, uint64_t *flags)
471 {
472 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
473 		*addr = adev->vm_manager.vram_base_offset + *addr -
474 			adev->gmc.vram_start;
475 	BUG_ON(*addr & 0xFFFF00000000003FULL);
476 
477 	if (!adev->gmc.translate_further)
478 		return;
479 
480 	if (level == AMDGPU_VM_PDB1) {
481 		/* Set the block fragment size */
482 		if (!(*flags & AMDGPU_PDE_PTE))
483 			*flags |= AMDGPU_PDE_BFS(0x9);
484 
485 	} else if (level == AMDGPU_VM_PDB0) {
486 		if (*flags & AMDGPU_PDE_PTE)
487 			*flags &= ~AMDGPU_PDE_PTE;
488 		else
489 			*flags |= AMDGPU_PTE_TF;
490 	}
491 }
492 
493 static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
494 				 struct amdgpu_bo_va_mapping *mapping,
495 				 uint64_t *flags)
496 {
497 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
498 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
499 
500 	*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
501 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
502 
503 	*flags &= ~AMDGPU_PTE_NOALLOC;
504 	*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
505 
506 	if (mapping->flags & AMDGPU_PTE_PRT) {
507 		*flags |= AMDGPU_PTE_PRT;
508 		*flags |= AMDGPU_PTE_SNOOPED;
509 		*flags |= AMDGPU_PTE_LOG;
510 		*flags |= AMDGPU_PTE_SYSTEM;
511 		*flags &= ~AMDGPU_PTE_VALID;
512 	}
513 }
514 
515 static unsigned gmc_v11_0_get_vbios_fb_size(struct amdgpu_device *adev)
516 {
517 	return 0;
518 }
519 
520 static const struct amdgpu_gmc_funcs gmc_v11_0_gmc_funcs = {
521 	.flush_gpu_tlb = gmc_v11_0_flush_gpu_tlb,
522 	.flush_gpu_tlb_pasid = gmc_v11_0_flush_gpu_tlb_pasid,
523 	.emit_flush_gpu_tlb = gmc_v11_0_emit_flush_gpu_tlb,
524 	.emit_pasid_mapping = gmc_v11_0_emit_pasid_mapping,
525 	.map_mtype = gmc_v11_0_map_mtype,
526 	.get_vm_pde = gmc_v11_0_get_vm_pde,
527 	.get_vm_pte = gmc_v11_0_get_vm_pte,
528 	.get_vbios_fb_size = gmc_v11_0_get_vbios_fb_size,
529 };
530 
531 static void gmc_v11_0_set_gmc_funcs(struct amdgpu_device *adev)
532 {
533 	adev->gmc.gmc_funcs = &gmc_v11_0_gmc_funcs;
534 }
535 
536 static void gmc_v11_0_set_umc_funcs(struct amdgpu_device *adev)
537 {
538 	switch (adev->ip_versions[UMC_HWIP][0]) {
539 	case IP_VERSION(8, 10, 0):
540 	case IP_VERSION(8, 11, 0):
541 		break;
542 	default:
543 		break;
544 	}
545 }
546 
547 
548 static void gmc_v11_0_set_mmhub_funcs(struct amdgpu_device *adev)
549 {
550 	switch (adev->ip_versions[MMHUB_HWIP][0]) {
551 	case IP_VERSION(3, 0, 2):
552 		adev->mmhub.funcs = &mmhub_v3_0_2_funcs;
553 		break;
554 	default:
555 		adev->mmhub.funcs = &mmhub_v3_0_funcs;
556 		break;
557 	}
558 }
559 
560 static void gmc_v11_0_set_gfxhub_funcs(struct amdgpu_device *adev)
561 {
562 	adev->gfxhub.funcs = &gfxhub_v3_0_funcs;
563 }
564 
565 static int gmc_v11_0_early_init(void *handle)
566 {
567 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
568 
569 	gmc_v11_0_set_gfxhub_funcs(adev);
570 	gmc_v11_0_set_mmhub_funcs(adev);
571 	gmc_v11_0_set_gmc_funcs(adev);
572 	gmc_v11_0_set_irq_funcs(adev);
573 	gmc_v11_0_set_umc_funcs(adev);
574 
575 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
576 	adev->gmc.shared_aperture_end =
577 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
578 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
579 	adev->gmc.private_aperture_end =
580 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
581 
582 	return 0;
583 }
584 
585 static int gmc_v11_0_late_init(void *handle)
586 {
587 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
588 	int r;
589 
590 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
591 	if (r)
592 		return r;
593 
594 	r = amdgpu_gmc_ras_late_init(adev);
595 	if (r)
596 		return r;
597 
598 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
599 }
600 
601 static void gmc_v11_0_vram_gtt_location(struct amdgpu_device *adev,
602 					struct amdgpu_gmc *mc)
603 {
604 	u64 base = 0;
605 
606 	base = adev->mmhub.funcs->get_fb_location(adev);
607 
608 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
609 	amdgpu_gmc_gart_location(adev, mc);
610 
611 	/* base offset of vram pages */
612 	adev->vm_manager.vram_base_offset = adev->mmhub.funcs->get_mc_fb_offset(adev);
613 }
614 
615 /**
616  * gmc_v11_0_mc_init - initialize the memory controller driver params
617  *
618  * @adev: amdgpu_device pointer
619  *
620  * Look up the amount of vram, vram width, and decide how to place
621  * vram and gart within the GPU's physical address space.
622  * Returns 0 for success.
623  */
624 static int gmc_v11_0_mc_init(struct amdgpu_device *adev)
625 {
626 	int r;
627 
628 	/* size in MB on si */
629 	adev->gmc.mc_vram_size =
630 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
631 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
632 
633 	if (!(adev->flags & AMD_IS_APU)) {
634 		r = amdgpu_device_resize_fb_bar(adev);
635 		if (r)
636 			return r;
637 	}
638 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
639 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
640 
641 	/* In case the PCI BAR is larger than the actual amount of vram */
642 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
643 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
644 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
645 
646 	/* set the gart size */
647 	if (amdgpu_gart_size == -1) {
648 		adev->gmc.gart_size = 512ULL << 20;
649 	} else
650 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
651 
652 	gmc_v11_0_vram_gtt_location(adev, &adev->gmc);
653 
654 	return 0;
655 }
656 
657 static int gmc_v11_0_gart_init(struct amdgpu_device *adev)
658 {
659 	int r;
660 
661 	if (adev->gart.bo) {
662 		WARN(1, "PCIE GART already initialized\n");
663 		return 0;
664 	}
665 
666 	/* Initialize common gart structure */
667 	r = amdgpu_gart_init(adev);
668 	if (r)
669 		return r;
670 
671 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
672 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
673 				 AMDGPU_PTE_EXECUTABLE;
674 
675 	return amdgpu_gart_table_vram_alloc(adev);
676 }
677 
678 static int gmc_v11_0_sw_init(void *handle)
679 {
680 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
681 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
682 
683 	adev->mmhub.funcs->init(adev);
684 
685 	spin_lock_init(&adev->gmc.invalidate_lock);
686 
687 	r = amdgpu_atomfirmware_get_vram_info(adev,
688 					      &vram_width, &vram_type, &vram_vendor);
689 	adev->gmc.vram_width = vram_width;
690 
691 	adev->gmc.vram_type = vram_type;
692 	adev->gmc.vram_vendor = vram_vendor;
693 
694 	switch (adev->ip_versions[GC_HWIP][0]) {
695 	case IP_VERSION(11, 0, 0):
696 	case IP_VERSION(11, 0, 1):
697 	case IP_VERSION(11, 0, 2):
698 		adev->num_vmhubs = 2;
699 		/*
700 		 * To fulfill 4-level page support,
701 		 * vm size is 256TB (48bit), maximum size,
702 		 * block size 512 (9bit)
703 		 */
704 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
705 		break;
706 	default:
707 		break;
708 	}
709 
710 	/* This interrupt is VMC page fault.*/
711 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_VMC,
712 			      VMC_1_0__SRCID__VM_FAULT,
713 			      &adev->gmc.vm_fault);
714 
715 	if (r)
716 		return r;
717 
718 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
719 			      UTCL2_1_0__SRCID__FAULT,
720 			      &adev->gmc.vm_fault);
721 	if (r)
722 		return r;
723 
724 	if (!amdgpu_sriov_vf(adev)) {
725 		/* interrupt sent to DF. */
726 		r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_DF, 0,
727 				      &adev->gmc.ecc_irq);
728 		if (r)
729 			return r;
730 	}
731 
732 	/*
733 	 * Set the internal MC address mask This is the max address of the GPU's
734 	 * internal address space.
735 	 */
736 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
737 
738 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
739 	if (r) {
740 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
741 		return r;
742 	}
743 
744 	r = gmc_v11_0_mc_init(adev);
745 	if (r)
746 		return r;
747 
748 	amdgpu_gmc_get_vbios_allocations(adev);
749 
750 	/* Memory manager */
751 	r = amdgpu_bo_init(adev);
752 	if (r)
753 		return r;
754 
755 	r = gmc_v11_0_gart_init(adev);
756 	if (r)
757 		return r;
758 
759 	/*
760 	 * number of VMs
761 	 * VMID 0 is reserved for System
762 	 * amdgpu graphics/compute will use VMIDs 1-7
763 	 * amdkfd will use VMIDs 8-15
764 	 */
765 	adev->vm_manager.first_kfd_vmid = 8;
766 
767 	amdgpu_vm_manager_init(adev);
768 
769 	return 0;
770 }
771 
772 /**
773  * gmc_v11_0_gart_fini - vm fini callback
774  *
775  * @adev: amdgpu_device pointer
776  *
777  * Tears down the driver GART/VM setup (CIK).
778  */
779 static void gmc_v11_0_gart_fini(struct amdgpu_device *adev)
780 {
781 	amdgpu_gart_table_vram_free(adev);
782 }
783 
784 static int gmc_v11_0_sw_fini(void *handle)
785 {
786 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
787 
788 	amdgpu_vm_manager_fini(adev);
789 	gmc_v11_0_gart_fini(adev);
790 	amdgpu_gem_force_release(adev);
791 	amdgpu_bo_fini(adev);
792 
793 	return 0;
794 }
795 
796 static void gmc_v11_0_init_golden_registers(struct amdgpu_device *adev)
797 {
798 }
799 
800 /**
801  * gmc_v11_0_gart_enable - gart enable
802  *
803  * @adev: amdgpu_device pointer
804  */
805 static int gmc_v11_0_gart_enable(struct amdgpu_device *adev)
806 {
807 	int r;
808 	bool value;
809 
810 	if (adev->gart.bo == NULL) {
811 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
812 		return -EINVAL;
813 	}
814 
815 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
816 
817 	r = adev->mmhub.funcs->gart_enable(adev);
818 	if (r)
819 		return r;
820 
821 	/* Flush HDP after it is initialized */
822 	adev->hdp.funcs->flush_hdp(adev, NULL);
823 
824 	value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
825 		false : true;
826 
827 	adev->mmhub.funcs->set_fault_enable_default(adev, value);
828 	gmc_v11_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB_0, 0);
829 
830 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
831 		 (unsigned)(adev->gmc.gart_size >> 20),
832 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
833 
834 	return 0;
835 }
836 
837 static int gmc_v11_0_hw_init(void *handle)
838 {
839 	int r;
840 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
841 
842 	/* The sequence of these two function calls matters.*/
843 	gmc_v11_0_init_golden_registers(adev);
844 
845 	r = gmc_v11_0_gart_enable(adev);
846 	if (r)
847 		return r;
848 
849 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
850 		adev->umc.funcs->init_registers(adev);
851 
852 	return 0;
853 }
854 
855 /**
856  * gmc_v11_0_gart_disable - gart disable
857  *
858  * @adev: amdgpu_device pointer
859  *
860  * This disables all VM page table.
861  */
862 static void gmc_v11_0_gart_disable(struct amdgpu_device *adev)
863 {
864 	adev->mmhub.funcs->gart_disable(adev);
865 }
866 
867 static int gmc_v11_0_hw_fini(void *handle)
868 {
869 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
870 
871 	if (amdgpu_sriov_vf(adev)) {
872 		/* full access mode, so don't touch any GMC register */
873 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
874 		return 0;
875 	}
876 
877 	amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
878 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
879 	gmc_v11_0_gart_disable(adev);
880 
881 	return 0;
882 }
883 
884 static int gmc_v11_0_suspend(void *handle)
885 {
886 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
887 
888 	gmc_v11_0_hw_fini(adev);
889 
890 	return 0;
891 }
892 
893 static int gmc_v11_0_resume(void *handle)
894 {
895 	int r;
896 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
897 
898 	r = gmc_v11_0_hw_init(adev);
899 	if (r)
900 		return r;
901 
902 	amdgpu_vmid_reset_all(adev);
903 
904 	return 0;
905 }
906 
907 static bool gmc_v11_0_is_idle(void *handle)
908 {
909 	/* MC is always ready in GMC v11.*/
910 	return true;
911 }
912 
913 static int gmc_v11_0_wait_for_idle(void *handle)
914 {
915 	/* There is no need to wait for MC idle in GMC v11.*/
916 	return 0;
917 }
918 
919 static int gmc_v11_0_soft_reset(void *handle)
920 {
921 	return 0;
922 }
923 
924 static int gmc_v11_0_set_clockgating_state(void *handle,
925 					   enum amd_clockgating_state state)
926 {
927 	int r;
928 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
929 
930 	r = adev->mmhub.funcs->set_clockgating(adev, state);
931 	if (r)
932 		return r;
933 
934 	return athub_v3_0_set_clockgating(adev, state);
935 }
936 
937 static void gmc_v11_0_get_clockgating_state(void *handle, u64 *flags)
938 {
939 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
940 
941 	adev->mmhub.funcs->get_clockgating(adev, flags);
942 
943 	athub_v3_0_get_clockgating(adev, flags);
944 }
945 
946 static int gmc_v11_0_set_powergating_state(void *handle,
947 					   enum amd_powergating_state state)
948 {
949 	return 0;
950 }
951 
952 const struct amd_ip_funcs gmc_v11_0_ip_funcs = {
953 	.name = "gmc_v11_0",
954 	.early_init = gmc_v11_0_early_init,
955 	.sw_init = gmc_v11_0_sw_init,
956 	.hw_init = gmc_v11_0_hw_init,
957 	.late_init = gmc_v11_0_late_init,
958 	.sw_fini = gmc_v11_0_sw_fini,
959 	.hw_fini = gmc_v11_0_hw_fini,
960 	.suspend = gmc_v11_0_suspend,
961 	.resume = gmc_v11_0_resume,
962 	.is_idle = gmc_v11_0_is_idle,
963 	.wait_for_idle = gmc_v11_0_wait_for_idle,
964 	.soft_reset = gmc_v11_0_soft_reset,
965 	.set_clockgating_state = gmc_v11_0_set_clockgating_state,
966 	.set_powergating_state = gmc_v11_0_set_powergating_state,
967 	.get_clockgating_state = gmc_v11_0_get_clockgating_state,
968 };
969 
970 const struct amdgpu_ip_block_version gmc_v11_0_ip_block = {
971 	.type = AMD_IP_BLOCK_TYPE_GMC,
972 	.major = 11,
973 	.minor = 0,
974 	.rev = 0,
975 	.funcs = &gmc_v11_0_ip_funcs,
976 };
977