xref: /linux/drivers/gpu/drm/amd/amdgpu/gmc_v12_0.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1 /*
2  * Copyright 2023 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_v12_0.h"
31 #include "athub/athub_4_1_0_sh_mask.h"
32 #include "athub/athub_4_1_0_offset.h"
33 #include "oss/osssys_7_0_0_offset.h"
34 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
35 #include "soc24_enum.h"
36 #include "soc24.h"
37 #include "soc15d.h"
38 #include "soc15_common.h"
39 #include "nbif_v6_3_1.h"
40 #include "gfxhub_v12_0.h"
41 #include "mmhub_v4_1_0.h"
42 #include "athub_v4_1_0.h"
43 
44 
45 static int gmc_v12_0_ecc_interrupt_state(struct amdgpu_device *adev,
46 					 struct amdgpu_irq_src *src,
47 					 unsigned type,
48 					 enum amdgpu_interrupt_state state)
49 {
50 	return 0;
51 }
52 
53 static int gmc_v12_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_MMHUB0(0), false);
61 		/* GFX HUB */
62 		/* This works because this interrupt is only
63 		 * enabled at init/resume and disabled in
64 		 * fini/suspend, so the overall state doesn't
65 		 * change over the course of suspend/resume.
66 		 */
67 		if (!adev->in_s0ix)
68 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), false);
69 		break;
70 	case AMDGPU_IRQ_STATE_ENABLE:
71 		/* MM HUB */
72 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), true);
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), true);
81 		break;
82 	default:
83 		break;
84 	}
85 
86 	return 0;
87 }
88 
89 static int gmc_v12_0_process_interrupt(struct amdgpu_device *adev,
90 				       struct amdgpu_irq_src *source,
91 				       struct amdgpu_iv_entry *entry)
92 {
93 	struct amdgpu_vmhub *hub;
94 	uint32_t status = 0;
95 	u64 addr;
96 
97 	addr = (u64)entry->src_data[0] << 12;
98 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
99 
100 	if (entry->client_id == SOC21_IH_CLIENTID_VMC)
101 		hub = &adev->vmhub[AMDGPU_MMHUB0(0)];
102 	else
103 		hub = &adev->vmhub[AMDGPU_GFXHUB(0)];
104 
105 	if (!amdgpu_sriov_vf(adev)) {
106 		/*
107 		 * Issue a dummy read to wait for the status register to
108 		 * be updated to avoid reading an incorrect value due to
109 		 * the new fast GRBM interface.
110 		 */
111 		if (entry->vmid_src == AMDGPU_GFXHUB(0))
112 			RREG32(hub->vm_l2_pro_fault_status);
113 
114 		status = RREG32(hub->vm_l2_pro_fault_status);
115 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
116 
117 		amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status,
118 					     entry->vmid_src ? AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0));
119 	}
120 
121 	if (printk_ratelimit()) {
122 		struct amdgpu_task_info *task_info;
123 
124 		dev_err(adev->dev,
125 			"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n",
126 			entry->vmid_src ? "mmhub" : "gfxhub",
127 			entry->src_id, entry->ring_id, entry->vmid, entry->pasid);
128 		task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid);
129 		if (task_info) {
130 			dev_err(adev->dev,
131 				" in process %s pid %d thread %s pid %d)\n",
132 				task_info->process_name, task_info->tgid,
133 				task_info->task_name, task_info->pid);
134 			amdgpu_vm_put_task_info(task_info);
135 		}
136 
137 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
138 				addr, entry->client_id);
139 
140 		if (!amdgpu_sriov_vf(adev))
141 			hub->vmhub_funcs->print_l2_protection_fault_status(adev, status);
142 	}
143 
144 	return 0;
145 }
146 
147 static const struct amdgpu_irq_src_funcs gmc_v12_0_irq_funcs = {
148 	.set = gmc_v12_0_vm_fault_interrupt_state,
149 	.process = gmc_v12_0_process_interrupt,
150 };
151 
152 static const struct amdgpu_irq_src_funcs gmc_v12_0_ecc_funcs = {
153 	.set = gmc_v12_0_ecc_interrupt_state,
154 	.process = amdgpu_umc_process_ecc_irq,
155 };
156 
157 static void gmc_v12_0_set_irq_funcs(struct amdgpu_device *adev)
158 {
159 	adev->gmc.vm_fault.num_types = 1;
160 	adev->gmc.vm_fault.funcs = &gmc_v12_0_irq_funcs;
161 
162 	if (!amdgpu_sriov_vf(adev)) {
163 		adev->gmc.ecc_irq.num_types = 1;
164 		adev->gmc.ecc_irq.funcs = &gmc_v12_0_ecc_funcs;
165 	}
166 }
167 
168 /**
169  * gmc_v12_0_use_invalidate_semaphore - judge whether to use semaphore
170  *
171  * @adev: amdgpu_device pointer
172  * @vmhub: vmhub type
173  *
174  */
175 static bool gmc_v12_0_use_invalidate_semaphore(struct amdgpu_device *adev,
176 				       uint32_t vmhub)
177 {
178 	return ((vmhub == AMDGPU_MMHUB0(0)) &&
179 		(!amdgpu_sriov_vf(adev)));
180 }
181 
182 static bool gmc_v12_0_get_vmid_pasid_mapping_info(
183 					struct amdgpu_device *adev,
184 					uint8_t vmid, uint16_t *p_pasid)
185 {
186 	*p_pasid = RREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid) & 0xffff;
187 
188 	return !!(*p_pasid);
189 }
190 
191 /*
192  * GART
193  * VMID 0 is the physical GPU addresses as used by the kernel.
194  * VMIDs 1-15 are used for userspace clients and are handled
195  * by the amdgpu vm/hsa code.
196  */
197 
198 static void gmc_v12_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
199 				   unsigned int vmhub, uint32_t flush_type)
200 {
201 	bool use_semaphore = gmc_v12_0_use_invalidate_semaphore(adev, vmhub);
202 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
203 	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
204 	u32 tmp;
205 	/* Use register 17 for GART */
206 	const unsigned eng = 17;
207 	unsigned int i;
208 	unsigned char hub_ip = 0;
209 
210 	hub_ip = (vmhub == AMDGPU_GFXHUB(0)) ?
211 		   GC_HWIP : MMHUB_HWIP;
212 
213 	spin_lock(&adev->gmc.invalidate_lock);
214 	/*
215 	 * It may lose gpuvm invalidate acknowldege state across power-gating
216 	 * off cycle, add semaphore acquire before invalidation and semaphore
217 	 * release after invalidation to avoid entering power gated state
218 	 * to WA the Issue
219 	 */
220 
221 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
222 	if (use_semaphore) {
223 		for (i = 0; i < adev->usec_timeout; i++) {
224 			/* a read return value of 1 means semaphore acuqire */
225 			tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
226 					    hub->eng_distance * eng, hub_ip);
227 			if (tmp & 0x1)
228 				break;
229 			udelay(1);
230 		}
231 
232 		if (i >= adev->usec_timeout)
233 			dev_err(adev->dev,
234 				"Timeout waiting for sem acquire in VM flush!\n");
235 	}
236 
237 	WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req, hub_ip);
238 
239 	/* Wait for ACK with a delay.*/
240 	for (i = 0; i < adev->usec_timeout; i++) {
241 		tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack +
242 				    hub->eng_distance * eng, hub_ip);
243 		tmp &= 1 << vmid;
244 		if (tmp)
245 			break;
246 
247 		udelay(1);
248 	}
249 
250 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
251 	if (use_semaphore)
252 		/*
253 		 * add semaphore release after invalidation,
254 		 * write with 0 means semaphore release
255 		 */
256 		WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
257 			      hub->eng_distance * eng, 0, hub_ip);
258 
259 	/* Issue additional private vm invalidation to MMHUB */
260 	if ((vmhub != AMDGPU_GFXHUB(0)) &&
261 	    (hub->vm_l2_bank_select_reserved_cid2) &&
262 		!amdgpu_sriov_vf(adev)) {
263 		inv_req = RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
264 		/* bit 25: RSERVED_CACHE_PRIVATE_INVALIDATION */
265 		inv_req |= (1 << 25);
266 		/* Issue private invalidation */
267 		WREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2, inv_req);
268 		/* Read back to ensure invalidation is done*/
269 		RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
270 	}
271 
272 	spin_unlock(&adev->gmc.invalidate_lock);
273 
274 	if (i < adev->usec_timeout)
275 		return;
276 
277 	dev_err(adev->dev, "Timeout waiting for VM flush ACK!\n");
278 }
279 
280 /**
281  * gmc_v12_0_flush_gpu_tlb - gart tlb flush callback
282  *
283  * @adev: amdgpu_device pointer
284  * @vmid: vm instance to flush
285  * @vmhub: which hub to flush
286  * @flush_type: the flush type
287  *
288  * Flush the TLB for the requested page table.
289  */
290 static void gmc_v12_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
291 					uint32_t vmhub, uint32_t flush_type)
292 {
293 	if ((vmhub == AMDGPU_GFXHUB(0)) && !adev->gfx.is_poweron)
294 		return;
295 
296 	/* flush hdp cache */
297 	adev->hdp.funcs->flush_hdp(adev, NULL);
298 
299 	/* This is necessary for SRIOV as well as for GFXOFF to function
300 	 * properly under bare metal
301 	 */
302 	if ((adev->gfx.kiq[0].ring.sched.ready || adev->mes.ring[0].sched.ready) &&
303 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) {
304 		struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
305 		const unsigned eng = 17;
306 		u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
307 		u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
308 		u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
309 
310 		amdgpu_gmc_fw_reg_write_reg_wait(adev, req, ack, inv_req,
311 				1 << vmid, GET_INST(GC, 0));
312 		return;
313 	}
314 
315 	mutex_lock(&adev->mman.gtt_window_lock);
316 	gmc_v12_0_flush_vm_hub(adev, vmid, vmhub, 0);
317 	mutex_unlock(&adev->mman.gtt_window_lock);
318 	return;
319 }
320 
321 /**
322  * gmc_v12_0_flush_gpu_tlb_pasid - tlb flush via pasid
323  *
324  * @adev: amdgpu_device pointer
325  * @pasid: pasid to be flush
326  * @flush_type: the flush type
327  * @all_hub: flush all hubs
328  * @inst: is used to select which instance of KIQ to use for the invalidation
329  *
330  * Flush the TLB for the requested pasid.
331  */
332 static void gmc_v12_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
333 					  uint16_t pasid, uint32_t flush_type,
334 					  bool all_hub, uint32_t inst)
335 {
336 	uint16_t queried;
337 	int vmid, i;
338 
339 	for (vmid = 1; vmid < 16; vmid++) {
340 		bool valid;
341 
342 		valid = gmc_v12_0_get_vmid_pasid_mapping_info(adev, vmid,
343 							      &queried);
344 		if (!valid || queried != pasid)
345 			continue;
346 
347 		if (all_hub) {
348 			for_each_set_bit(i, adev->vmhubs_mask,
349 					 AMDGPU_MAX_VMHUBS)
350 				gmc_v12_0_flush_gpu_tlb(adev, vmid, i,
351 							flush_type);
352 		} else {
353 			gmc_v12_0_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0),
354 						flush_type);
355 		}
356 	}
357 }
358 
359 static uint64_t gmc_v12_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
360 					     unsigned vmid, uint64_t pd_addr)
361 {
362 	bool use_semaphore = gmc_v12_0_use_invalidate_semaphore(ring->adev, ring->vm_hub);
363 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
364 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
365 	unsigned eng = ring->vm_inv_eng;
366 
367 	/*
368 	 * It may lose gpuvm invalidate acknowldege state across power-gating
369 	 * off cycle, add semaphore acquire before invalidation and semaphore
370 	 * release after invalidation to avoid entering power gated state
371 	 * to WA the Issue
372 	 */
373 
374 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
375 	if (use_semaphore)
376 		/* a read return value of 1 means semaphore acuqire */
377 		amdgpu_ring_emit_reg_wait(ring,
378 					  hub->vm_inv_eng0_sem +
379 					  hub->eng_distance * eng, 0x1, 0x1);
380 
381 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
382 			      (hub->ctx_addr_distance * vmid),
383 			      lower_32_bits(pd_addr));
384 
385 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
386 			      (hub->ctx_addr_distance * vmid),
387 			      upper_32_bits(pd_addr));
388 
389 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
390 					    hub->eng_distance * eng,
391 					    hub->vm_inv_eng0_ack +
392 					    hub->eng_distance * eng,
393 					    req, 1 << vmid);
394 
395 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
396 	if (use_semaphore)
397 		/*
398 		 * add semaphore release after invalidation,
399 		 * write with 0 means semaphore release
400 		 */
401 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
402 				      hub->eng_distance * eng, 0);
403 
404 	return pd_addr;
405 }
406 
407 static void gmc_v12_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
408 					 unsigned pasid)
409 {
410 	struct amdgpu_device *adev = ring->adev;
411 	uint32_t reg;
412 
413 	/* MES fw manages IH_VMID_x_LUT updating */
414 	if (ring->is_mes_queue)
415 		return;
416 
417 	if (ring->vm_hub == AMDGPU_GFXHUB(0))
418 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid;
419 	else
420 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT_MM) + vmid;
421 
422 	amdgpu_ring_emit_wreg(ring, reg, pasid);
423 }
424 
425 /*
426  * PTE format:
427  * 63 P
428  * 62:59 reserved
429  * 58 D
430  * 57 G
431  * 56 T
432  * 55:54 M
433  * 53:52 SW
434  * 51:48 reserved for future
435  * 47:12 4k physical page base address
436  * 11:7 fragment
437  * 6 write
438  * 5 read
439  * 4 exe
440  * 3 Z
441  * 2 snooped
442  * 1 system
443  * 0 valid
444  *
445  * PDE format:
446  * 63 P
447  * 62:58 block fragment size
448  * 57 reserved
449  * 56 A
450  * 55:54 M
451  * 53:52 reserved
452  * 51:48 reserved for future
453  * 47:6 physical base address of PD or PTE
454  * 5:3 reserved
455  * 2 C
456  * 1 system
457  * 0 valid
458  */
459 
460 static uint64_t gmc_v12_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
461 {
462 	switch (flags) {
463 	case AMDGPU_VM_MTYPE_DEFAULT:
464 		return AMDGPU_PTE_MTYPE_GFX12(0ULL, MTYPE_NC);
465 	case AMDGPU_VM_MTYPE_NC:
466 		return AMDGPU_PTE_MTYPE_GFX12(0ULL, MTYPE_NC);
467 	case AMDGPU_VM_MTYPE_UC:
468 		return AMDGPU_PTE_MTYPE_GFX12(0ULL, MTYPE_UC);
469 	default:
470 		return AMDGPU_PTE_MTYPE_GFX12(0ULL, MTYPE_NC);
471 	}
472 }
473 
474 static void gmc_v12_0_get_vm_pde(struct amdgpu_device *adev, int level,
475 				 uint64_t *addr, uint64_t *flags)
476 {
477 	if (!(*flags & AMDGPU_PDE_PTE_GFX12) && !(*flags & AMDGPU_PTE_SYSTEM))
478 		*addr = adev->vm_manager.vram_base_offset + *addr -
479 			adev->gmc.vram_start;
480 	BUG_ON(*addr & 0xFFFF00000000003FULL);
481 
482 	if (!adev->gmc.translate_further)
483 		return;
484 
485 	if (level == AMDGPU_VM_PDB1) {
486 		/* Set the block fragment size */
487 		if (!(*flags & AMDGPU_PDE_PTE_GFX12))
488 			*flags |= AMDGPU_PDE_BFS_GFX12(0x9);
489 
490 	} else if (level == AMDGPU_VM_PDB0) {
491 		if (*flags & AMDGPU_PDE_PTE_GFX12)
492 			*flags &= ~AMDGPU_PDE_PTE_GFX12;
493 	}
494 }
495 
496 static void gmc_v12_0_get_vm_pte(struct amdgpu_device *adev,
497 				 struct amdgpu_bo_va_mapping *mapping,
498 				 uint64_t *flags)
499 {
500 	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
501 	struct amdgpu_device *bo_adev;
502 	bool coherent, is_system;
503 
504 
505 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
506 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
507 
508 	*flags &= ~AMDGPU_PTE_MTYPE_GFX12_MASK;
509 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_GFX12_MASK);
510 
511 	if (mapping->flags & AMDGPU_PTE_PRT_GFX12) {
512 		*flags |= AMDGPU_PTE_PRT_GFX12;
513 		*flags |= AMDGPU_PTE_SNOOPED;
514 		*flags |= AMDGPU_PTE_SYSTEM;
515 		*flags |= AMDGPU_PTE_IS_PTE;
516 		*flags &= ~AMDGPU_PTE_VALID;
517 	}
518 
519 	if (!bo)
520 		return;
521 
522 	if (bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
523 			       AMDGPU_GEM_CREATE_UNCACHED))
524 		*flags = AMDGPU_PTE_MTYPE_GFX12(*flags, MTYPE_UC);
525 
526 	bo_adev = amdgpu_ttm_adev(bo->tbo.bdev);
527 	coherent = bo->flags & AMDGPU_GEM_CREATE_COHERENT;
528 	is_system = (bo->tbo.resource->mem_type == TTM_PL_TT) ||
529 		(bo->tbo.resource->mem_type == AMDGPU_PL_PREEMPT);
530 
531 	if (bo && bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC)
532 		*flags |= AMDGPU_PTE_DCC;
533 
534 	/* WA for HW bug */
535 	if (is_system || ((bo_adev != adev) && coherent))
536 		*flags = AMDGPU_PTE_MTYPE_GFX12(*flags, MTYPE_NC);
537 
538 }
539 
540 static unsigned gmc_v12_0_get_vbios_fb_size(struct amdgpu_device *adev)
541 {
542 	return 0;
543 }
544 
545 static unsigned int gmc_v12_0_get_dcc_alignment(struct amdgpu_device *adev)
546 {
547 	unsigned int max_tex_channel_caches, alignment;
548 
549 	if (amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(12, 0, 0) &&
550 	    amdgpu_ip_version(adev, GC_HWIP, 0) != IP_VERSION(12, 0, 1))
551 		return 0;
552 
553 	max_tex_channel_caches = adev->gfx.config.max_texture_channel_caches;
554 	if (is_power_of_2(max_tex_channel_caches))
555 		alignment = (unsigned int)(max_tex_channel_caches / SZ_4);
556 	else
557 		alignment = roundup_pow_of_two(max_tex_channel_caches);
558 
559 	return (unsigned int)(alignment * max_tex_channel_caches * SZ_1K);
560 }
561 
562 static const struct amdgpu_gmc_funcs gmc_v12_0_gmc_funcs = {
563 	.flush_gpu_tlb = gmc_v12_0_flush_gpu_tlb,
564 	.flush_gpu_tlb_pasid = gmc_v12_0_flush_gpu_tlb_pasid,
565 	.emit_flush_gpu_tlb = gmc_v12_0_emit_flush_gpu_tlb,
566 	.emit_pasid_mapping = gmc_v12_0_emit_pasid_mapping,
567 	.map_mtype = gmc_v12_0_map_mtype,
568 	.get_vm_pde = gmc_v12_0_get_vm_pde,
569 	.get_vm_pte = gmc_v12_0_get_vm_pte,
570 	.get_vbios_fb_size = gmc_v12_0_get_vbios_fb_size,
571 	.get_dcc_alignment = gmc_v12_0_get_dcc_alignment,
572 };
573 
574 static void gmc_v12_0_set_gmc_funcs(struct amdgpu_device *adev)
575 {
576 	adev->gmc.gmc_funcs = &gmc_v12_0_gmc_funcs;
577 }
578 
579 static void gmc_v12_0_set_umc_funcs(struct amdgpu_device *adev)
580 {
581 }
582 
583 
584 static void gmc_v12_0_set_mmhub_funcs(struct amdgpu_device *adev)
585 {
586 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
587 	case IP_VERSION(4, 1, 0):
588 		adev->mmhub.funcs = &mmhub_v4_1_0_funcs;
589 		break;
590 	default:
591 		break;
592 	}
593 }
594 
595 static void gmc_v12_0_set_gfxhub_funcs(struct amdgpu_device *adev)
596 {
597 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
598 	case IP_VERSION(12, 0, 0):
599 	case IP_VERSION(12, 0, 1):
600 		adev->gfxhub.funcs = &gfxhub_v12_0_funcs;
601 		break;
602 	default:
603 		break;
604 	}
605 }
606 
607 static int gmc_v12_0_early_init(void *handle)
608 {
609 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
610 
611 	gmc_v12_0_set_gfxhub_funcs(adev);
612 	gmc_v12_0_set_mmhub_funcs(adev);
613 	gmc_v12_0_set_gmc_funcs(adev);
614 	gmc_v12_0_set_irq_funcs(adev);
615 	gmc_v12_0_set_umc_funcs(adev);
616 
617 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
618 	adev->gmc.shared_aperture_end =
619 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
620 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
621 	adev->gmc.private_aperture_end =
622 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
623 
624 	return 0;
625 }
626 
627 static int gmc_v12_0_late_init(void *handle)
628 {
629 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
630 	int r;
631 
632 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
633 	if (r)
634 		return r;
635 
636 	r = amdgpu_gmc_ras_late_init(adev);
637 	if (r)
638 		return r;
639 
640 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
641 }
642 
643 static void gmc_v12_0_vram_gtt_location(struct amdgpu_device *adev,
644 					struct amdgpu_gmc *mc)
645 {
646 	u64 base = 0;
647 
648 	base = adev->mmhub.funcs->get_fb_location(adev);
649 
650 	amdgpu_gmc_set_agp_default(adev, mc);
651 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
652 	amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_LOW);
653 	if (!amdgpu_sriov_vf(adev) && (amdgpu_agp == 1))
654 		amdgpu_gmc_agp_location(adev, mc);
655 
656 	/* base offset of vram pages */
657 	if (amdgpu_sriov_vf(adev))
658 		adev->vm_manager.vram_base_offset = 0;
659 	else
660 		adev->vm_manager.vram_base_offset = adev->mmhub.funcs->get_mc_fb_offset(adev);
661 }
662 
663 /**
664  * gmc_v12_0_mc_init - initialize the memory controller driver params
665  *
666  * @adev: amdgpu_device pointer
667  *
668  * Look up the amount of vram, vram width, and decide how to place
669  * vram and gart within the GPU's physical address space.
670  * Returns 0 for success.
671  */
672 static int gmc_v12_0_mc_init(struct amdgpu_device *adev)
673 {
674 	int r;
675 
676 	/* size in MB on si */
677 	adev->gmc.mc_vram_size =
678 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
679 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
680 
681 	if (!(adev->flags & AMD_IS_APU)) {
682 		r = amdgpu_device_resize_fb_bar(adev);
683 		if (r)
684 			return r;
685 	}
686 
687 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
688 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
689 
690 #ifdef CONFIG_X86_64
691 	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) {
692 		adev->gmc.aper_base = adev->mmhub.funcs->get_mc_fb_offset(adev);
693 		adev->gmc.aper_size = adev->gmc.real_vram_size;
694 	}
695 #endif
696 	/* In case the PCI BAR is larger than the actual amount of vram */
697 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
698 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
699 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
700 
701 	/* set the gart size */
702 	if (amdgpu_gart_size == -1) {
703 		adev->gmc.gart_size = 512ULL << 20;
704 	} else
705 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
706 
707 	gmc_v12_0_vram_gtt_location(adev, &adev->gmc);
708 
709 	return 0;
710 }
711 
712 static int gmc_v12_0_gart_init(struct amdgpu_device *adev)
713 {
714 	int r;
715 
716 	if (adev->gart.bo) {
717 		WARN(1, "PCIE GART already initialized\n");
718 		return 0;
719 	}
720 
721 	/* Initialize common gart structure */
722 	r = amdgpu_gart_init(adev);
723 	if (r)
724 		return r;
725 
726 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
727 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_GFX12(0ULL, MTYPE_UC) |
728 				    AMDGPU_PTE_EXECUTABLE |
729 				    AMDGPU_PTE_IS_PTE;
730 
731 	return amdgpu_gart_table_vram_alloc(adev);
732 }
733 
734 static int gmc_v12_0_sw_init(void *handle)
735 {
736 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
737 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
738 
739 	adev->mmhub.funcs->init(adev);
740 
741 	adev->gfxhub.funcs->init(adev);
742 
743 	spin_lock_init(&adev->gmc.invalidate_lock);
744 
745 	r = amdgpu_atomfirmware_get_vram_info(adev,
746 					      &vram_width, &vram_type, &vram_vendor);
747 	adev->gmc.vram_width = vram_width;
748 
749 	adev->gmc.vram_type = vram_type;
750 	adev->gmc.vram_vendor = vram_vendor;
751 
752 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
753 	case IP_VERSION(12, 0, 0):
754 	case IP_VERSION(12, 0, 1):
755 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
756 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
757 		/*
758 		 * To fulfill 4-level page support,
759 		 * vm size is 256TB (48bit), maximum size,
760 		 * block size 512 (9bit)
761 		 */
762 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
763 		break;
764 	default:
765 		break;
766 	}
767 
768 	/* This interrupt is VMC page fault.*/
769 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_VMC,
770 			      VMC_1_0__SRCID__VM_FAULT,
771 			      &adev->gmc.vm_fault);
772 
773 	if (r)
774 		return r;
775 
776 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
777 			      UTCL2_1_0__SRCID__FAULT,
778 			      &adev->gmc.vm_fault);
779 	if (r)
780 		return r;
781 
782 	if (!amdgpu_sriov_vf(adev)) {
783 		/* interrupt sent to DF. */
784 		r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_DF, 0,
785 				      &adev->gmc.ecc_irq);
786 		if (r)
787 			return r;
788 	}
789 
790 	/*
791 	 * Set the internal MC address mask This is the max address of the GPU's
792 	 * internal address space.
793 	 */
794 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
795 
796 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
797 	if (r) {
798 		printk(KERN_WARNING "amdgpu: No suitable DMA available.\n");
799 		return r;
800 	}
801 
802 	adev->need_swiotlb = drm_need_swiotlb(44);
803 
804 	r = gmc_v12_0_mc_init(adev);
805 	if (r)
806 		return r;
807 
808 	amdgpu_gmc_get_vbios_allocations(adev);
809 
810 	/* Memory manager */
811 	r = amdgpu_bo_init(adev);
812 	if (r)
813 		return r;
814 
815 	r = gmc_v12_0_gart_init(adev);
816 	if (r)
817 		return r;
818 
819 	/*
820 	 * number of VMs
821 	 * VMID 0 is reserved for System
822 	 * amdgpu graphics/compute will use VMIDs 1-7
823 	 * amdkfd will use VMIDs 8-15
824 	 */
825 	adev->vm_manager.first_kfd_vmid = 8;
826 
827 	amdgpu_vm_manager_init(adev);
828 
829 	return 0;
830 }
831 
832 /**
833  * gmc_v12_0_gart_fini - vm fini callback
834  *
835  * @adev: amdgpu_device pointer
836  *
837  * Tears down the driver GART/VM setup (CIK).
838  */
839 static void gmc_v12_0_gart_fini(struct amdgpu_device *adev)
840 {
841 	amdgpu_gart_table_vram_free(adev);
842 }
843 
844 static int gmc_v12_0_sw_fini(void *handle)
845 {
846 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
847 
848 	amdgpu_vm_manager_fini(adev);
849 	gmc_v12_0_gart_fini(adev);
850 	amdgpu_gem_force_release(adev);
851 	amdgpu_bo_fini(adev);
852 
853 	return 0;
854 }
855 
856 static void gmc_v12_0_init_golden_registers(struct amdgpu_device *adev)
857 {
858 }
859 
860 /**
861  * gmc_v12_0_gart_enable - gart enable
862  *
863  * @adev: amdgpu_device pointer
864  */
865 static int gmc_v12_0_gart_enable(struct amdgpu_device *adev)
866 {
867 	int r;
868 	bool value;
869 
870 	if (adev->gart.bo == NULL) {
871 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
872 		return -EINVAL;
873 	}
874 
875 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
876 
877 	r = adev->mmhub.funcs->gart_enable(adev);
878 	if (r)
879 		return r;
880 
881 	/* Flush HDP after it is initialized */
882 	adev->hdp.funcs->flush_hdp(adev, NULL);
883 
884 	value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
885 		false : true;
886 
887 	adev->mmhub.funcs->set_fault_enable_default(adev, value);
888 	gmc_v12_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB0(0), 0);
889 
890 	dev_info(adev->dev, "PCIE GART of %uM enabled (table at 0x%016llX).\n",
891 		 (unsigned)(adev->gmc.gart_size >> 20),
892 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
893 
894 	return 0;
895 }
896 
897 static int gmc_v12_0_hw_init(void *handle)
898 {
899 	int r;
900 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
901 
902 	/* The sequence of these two function calls matters.*/
903 	gmc_v12_0_init_golden_registers(adev);
904 
905 	r = gmc_v12_0_gart_enable(adev);
906 	if (r)
907 		return r;
908 
909 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
910 		adev->umc.funcs->init_registers(adev);
911 
912 	return 0;
913 }
914 
915 /**
916  * gmc_v12_0_gart_disable - gart disable
917  *
918  * @adev: amdgpu_device pointer
919  *
920  * This disables all VM page table.
921  */
922 static void gmc_v12_0_gart_disable(struct amdgpu_device *adev)
923 {
924 	adev->mmhub.funcs->gart_disable(adev);
925 }
926 
927 static int gmc_v12_0_hw_fini(void *handle)
928 {
929 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
930 
931 	if (amdgpu_sriov_vf(adev)) {
932 		/* full access mode, so don't touch any GMC register */
933 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
934 		return 0;
935 	}
936 
937 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
938 
939 	if (adev->gmc.ecc_irq.funcs &&
940 		amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC))
941 		amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
942 
943 	gmc_v12_0_gart_disable(adev);
944 
945 	return 0;
946 }
947 
948 static int gmc_v12_0_suspend(void *handle)
949 {
950 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
951 
952 	gmc_v12_0_hw_fini(adev);
953 
954 	return 0;
955 }
956 
957 static int gmc_v12_0_resume(void *handle)
958 {
959 	int r;
960 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
961 
962 	r = gmc_v12_0_hw_init(adev);
963 	if (r)
964 		return r;
965 
966 	amdgpu_vmid_reset_all(adev);
967 
968 	return 0;
969 }
970 
971 static bool gmc_v12_0_is_idle(void *handle)
972 {
973 	/* MC is always ready in GMC v11.*/
974 	return true;
975 }
976 
977 static int gmc_v12_0_wait_for_idle(void *handle)
978 {
979 	/* There is no need to wait for MC idle in GMC v11.*/
980 	return 0;
981 }
982 
983 static int gmc_v12_0_soft_reset(void *handle)
984 {
985 	return 0;
986 }
987 
988 static int gmc_v12_0_set_clockgating_state(void *handle,
989 					   enum amd_clockgating_state state)
990 {
991 	int r;
992 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
993 
994 	r = adev->mmhub.funcs->set_clockgating(adev, state);
995 	if (r)
996 		return r;
997 
998 	return athub_v4_1_0_set_clockgating(adev, state);
999 }
1000 
1001 static void gmc_v12_0_get_clockgating_state(void *handle, u64 *flags)
1002 {
1003 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1004 
1005 	adev->mmhub.funcs->get_clockgating(adev, flags);
1006 
1007 	athub_v4_1_0_get_clockgating(adev, flags);
1008 }
1009 
1010 static int gmc_v12_0_set_powergating_state(void *handle,
1011 					   enum amd_powergating_state state)
1012 {
1013 	return 0;
1014 }
1015 
1016 const struct amd_ip_funcs gmc_v12_0_ip_funcs = {
1017 	.name = "gmc_v12_0",
1018 	.early_init = gmc_v12_0_early_init,
1019 	.sw_init = gmc_v12_0_sw_init,
1020 	.hw_init = gmc_v12_0_hw_init,
1021 	.late_init = gmc_v12_0_late_init,
1022 	.sw_fini = gmc_v12_0_sw_fini,
1023 	.hw_fini = gmc_v12_0_hw_fini,
1024 	.suspend = gmc_v12_0_suspend,
1025 	.resume = gmc_v12_0_resume,
1026 	.is_idle = gmc_v12_0_is_idle,
1027 	.wait_for_idle = gmc_v12_0_wait_for_idle,
1028 	.soft_reset = gmc_v12_0_soft_reset,
1029 	.set_clockgating_state = gmc_v12_0_set_clockgating_state,
1030 	.set_powergating_state = gmc_v12_0_set_powergating_state,
1031 	.get_clockgating_state = gmc_v12_0_get_clockgating_state,
1032 };
1033 
1034 const struct amdgpu_ip_block_version gmc_v12_0_ip_block = {
1035 	.type = AMD_IP_BLOCK_TYPE_GMC,
1036 	.major = 12,
1037 	.minor = 0,
1038 	.rev = 0,
1039 	.funcs = &gmc_v12_0_ip_funcs,
1040 };
1041