xref: /linux/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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 
26 #include <drm/drm_cache.h>
27 
28 #include "amdgpu.h"
29 #include "amdgpu_atomfirmware.h"
30 #include "gmc_v11_0.h"
31 #include "umc_v8_10.h"
32 #include "athub/athub_3_0_0_sh_mask.h"
33 #include "athub/athub_3_0_0_offset.h"
34 #include "dcn/dcn_3_2_0_offset.h"
35 #include "dcn/dcn_3_2_0_sh_mask.h"
36 #include "oss/osssys_6_0_0_offset.h"
37 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
38 #include "navi10_enum.h"
39 #include "soc15.h"
40 #include "soc15d.h"
41 #include "soc15_common.h"
42 #include "nbio_v4_3.h"
43 #include "gfxhub_v3_0.h"
44 #include "gfxhub_v3_0_3.h"
45 #include "gfxhub_v11_5_0.h"
46 #include "mmhub_v3_0.h"
47 #include "mmhub_v3_0_1.h"
48 #include "mmhub_v3_0_2.h"
49 #include "mmhub_v3_3.h"
50 #include "athub_v3_0.h"
51 
52 
53 static int gmc_v11_0_ecc_interrupt_state(struct amdgpu_device *adev,
54 					 struct amdgpu_irq_src *src,
55 					 unsigned int type,
56 					 enum amdgpu_interrupt_state state)
57 {
58 	return 0;
59 }
60 
61 static int
62 gmc_v11_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
63 				   struct amdgpu_irq_src *src, unsigned int type,
64 				   enum amdgpu_interrupt_state state)
65 {
66 	switch (state) {
67 	case AMDGPU_IRQ_STATE_DISABLE:
68 		/* MM HUB */
69 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), false);
70 		/* GFX HUB */
71 		/* This works because this interrupt is only
72 		 * enabled at init/resume and disabled in
73 		 * fini/suspend, so the overall state doesn't
74 		 * change over the course of suspend/resume.
75 		 */
76 		if (!adev->in_s0ix && (adev->in_runpm || adev->in_suspend ||
77 							   amdgpu_in_reset(adev)))
78 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), false);
79 		break;
80 	case AMDGPU_IRQ_STATE_ENABLE:
81 		/* MM HUB */
82 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB0(0), true);
83 		/* GFX HUB */
84 		/* This works because this interrupt is only
85 		 * enabled at init/resume and disabled in
86 		 * fini/suspend, so the overall state doesn't
87 		 * change over the course of suspend/resume.
88 		 */
89 		if (!adev->in_s0ix)
90 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB(0), true);
91 		break;
92 	default:
93 		break;
94 	}
95 
96 	return 0;
97 }
98 
99 static int gmc_v11_0_process_interrupt(struct amdgpu_device *adev,
100 				       struct amdgpu_irq_src *source,
101 				       struct amdgpu_iv_entry *entry)
102 {
103 	uint32_t vmhub_index = entry->client_id == SOC21_IH_CLIENTID_VMC ?
104 			       AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0);
105 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub_index];
106 	uint32_t status = 0;
107 	u64 addr;
108 
109 	addr = (u64)entry->src_data[0] << 12;
110 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
111 
112 	if (!amdgpu_sriov_vf(adev)) {
113 		/*
114 		 * Issue a dummy read to wait for the status register to
115 		 * be updated to avoid reading an incorrect value due to
116 		 * the new fast GRBM interface.
117 		 */
118 		if (entry->vmid_src == AMDGPU_GFXHUB(0))
119 			RREG32(hub->vm_l2_pro_fault_status);
120 
121 		status = RREG32(hub->vm_l2_pro_fault_status);
122 		WREG32_P(hub->vm_l2_pro_fault_cntl, 1, ~1);
123 
124 		amdgpu_vm_update_fault_cache(adev, entry->pasid, addr, status,
125 					     entry->vmid_src ? AMDGPU_MMHUB0(0) : AMDGPU_GFXHUB(0));
126 	}
127 
128 	if (printk_ratelimit()) {
129 		struct amdgpu_task_info *task_info;
130 
131 		dev_err(adev->dev,
132 			"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u)\n",
133 			entry->vmid_src ? "mmhub" : "gfxhub",
134 			entry->src_id, entry->ring_id, entry->vmid, entry->pasid);
135 		task_info = amdgpu_vm_get_task_info_pasid(adev, entry->pasid);
136 		if (task_info) {
137 			dev_err(adev->dev,
138 				" in process %s pid %d thread %s pid %d)\n",
139 				task_info->process_name, task_info->tgid,
140 				task_info->task_name, task_info->pid);
141 			amdgpu_vm_put_task_info(task_info);
142 		}
143 
144 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
145 				addr, entry->client_id);
146 
147 		/* Only print L2 fault status if the status register could be read and
148 		 * contains useful information
149 		 */
150 		if (status != 0)
151 			hub->vmhub_funcs->print_l2_protection_fault_status(adev, status);
152 	}
153 
154 	return 0;
155 }
156 
157 static const struct amdgpu_irq_src_funcs gmc_v11_0_irq_funcs = {
158 	.set = gmc_v11_0_vm_fault_interrupt_state,
159 	.process = gmc_v11_0_process_interrupt,
160 };
161 
162 static const struct amdgpu_irq_src_funcs gmc_v11_0_ecc_funcs = {
163 	.set = gmc_v11_0_ecc_interrupt_state,
164 	.process = amdgpu_umc_process_ecc_irq,
165 };
166 
167 static void gmc_v11_0_set_irq_funcs(struct amdgpu_device *adev)
168 {
169 	adev->gmc.vm_fault.num_types = 1;
170 	adev->gmc.vm_fault.funcs = &gmc_v11_0_irq_funcs;
171 
172 	if (!amdgpu_sriov_vf(adev)) {
173 		adev->gmc.ecc_irq.num_types = 1;
174 		adev->gmc.ecc_irq.funcs = &gmc_v11_0_ecc_funcs;
175 	}
176 }
177 
178 /**
179  * gmc_v11_0_use_invalidate_semaphore - judge whether to use semaphore
180  *
181  * @adev: amdgpu_device pointer
182  * @vmhub: vmhub type
183  *
184  */
185 static bool gmc_v11_0_use_invalidate_semaphore(struct amdgpu_device *adev,
186 				       uint32_t vmhub)
187 {
188 	return ((vmhub == AMDGPU_MMHUB0(0)) &&
189 		(!amdgpu_sriov_vf(adev)));
190 }
191 
192 static bool gmc_v11_0_get_vmid_pasid_mapping_info(
193 					struct amdgpu_device *adev,
194 					uint8_t vmid, uint16_t *p_pasid)
195 {
196 	*p_pasid = RREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid) & 0xffff;
197 
198 	return !!(*p_pasid);
199 }
200 
201 /**
202  * gmc_v11_0_flush_gpu_tlb - gart tlb flush callback
203  *
204  * @adev: amdgpu_device pointer
205  * @vmid: vm instance to flush
206  * @vmhub: which hub to flush
207  * @flush_type: the flush type
208  *
209  * Flush the TLB for the requested page table.
210  */
211 static void gmc_v11_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
212 					uint32_t vmhub, uint32_t flush_type)
213 {
214 	bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(adev, vmhub);
215 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
216 	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
217 	/* Use register 17 for GART */
218 	const unsigned int eng = 17;
219 	unsigned char hub_ip;
220 	u32 sem, req, ack;
221 	unsigned int i;
222 	u32 tmp;
223 
224 	if ((vmhub == AMDGPU_GFXHUB(0)) && !adev->gfx.is_poweron)
225 		return;
226 
227 	sem = hub->vm_inv_eng0_sem + hub->eng_distance * eng;
228 	req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
229 	ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
230 
231 	/* flush hdp cache */
232 	adev->hdp.funcs->flush_hdp(adev, NULL);
233 
234 	/* This is necessary for SRIOV as well as for GFXOFF to function
235 	 * properly under bare metal
236 	 */
237 	if ((adev->gfx.kiq[0].ring.sched.ready || adev->mes.ring[0].sched.ready) &&
238 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) {
239 		amdgpu_gmc_fw_reg_write_reg_wait(adev, req, ack, inv_req,
240 						 1 << vmid, GET_INST(GC, 0));
241 		return;
242 	}
243 
244 	/* This path is needed before KIQ/MES/GFXOFF are set up */
245 	hub_ip = (vmhub == AMDGPU_GFXHUB(0)) ? GC_HWIP : MMHUB_HWIP;
246 
247 	spin_lock(&adev->gmc.invalidate_lock);
248 	/*
249 	 * It may lose gpuvm invalidate acknowldege state across power-gating
250 	 * off cycle, add semaphore acquire before invalidation and semaphore
251 	 * release after invalidation to avoid entering power gated state
252 	 * to WA the Issue
253 	 */
254 
255 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
256 	if (use_semaphore) {
257 		for (i = 0; i < adev->usec_timeout; i++) {
258 			/* a read return value of 1 means semaphore acuqire */
259 			tmp = RREG32_RLC_NO_KIQ(sem, hub_ip);
260 			if (tmp & 0x1)
261 				break;
262 			udelay(1);
263 		}
264 
265 		if (i >= adev->usec_timeout)
266 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
267 	}
268 
269 	WREG32_RLC_NO_KIQ(req, inv_req, hub_ip);
270 
271 	/* Wait for ACK with a delay.*/
272 	for (i = 0; i < adev->usec_timeout; i++) {
273 		tmp = RREG32_RLC_NO_KIQ(ack, hub_ip);
274 		tmp &= 1 << vmid;
275 		if (tmp)
276 			break;
277 
278 		udelay(1);
279 	}
280 
281 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
282 	if (use_semaphore)
283 		WREG32_RLC_NO_KIQ(sem, 0, hub_ip);
284 
285 	/* Issue additional private vm invalidation to MMHUB */
286 	if ((vmhub != AMDGPU_GFXHUB(0)) &&
287 	    (hub->vm_l2_bank_select_reserved_cid2) &&
288 		!amdgpu_sriov_vf(adev)) {
289 		inv_req = RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
290 		/* bit 25: RSERVED_CACHE_PRIVATE_INVALIDATION */
291 		inv_req |= (1 << 25);
292 		/* Issue private invalidation */
293 		WREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2, inv_req);
294 		/* Read back to ensure invalidation is done*/
295 		RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
296 	}
297 
298 	spin_unlock(&adev->gmc.invalidate_lock);
299 
300 	if (i >= adev->usec_timeout)
301 		dev_err(adev->dev, "Timeout waiting for VM flush ACK!\n");
302 }
303 
304 /**
305  * gmc_v11_0_flush_gpu_tlb_pasid - tlb flush via pasid
306  *
307  * @adev: amdgpu_device pointer
308  * @pasid: pasid to be flush
309  * @flush_type: the flush type
310  * @all_hub: flush all hubs
311  * @inst: is used to select which instance of KIQ to use for the invalidation
312  *
313  * Flush the TLB for the requested pasid.
314  */
315 static void gmc_v11_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
316 					  uint16_t pasid, uint32_t flush_type,
317 					  bool all_hub, uint32_t inst)
318 {
319 	uint16_t queried;
320 	int vmid, i;
321 
322 	for (vmid = 1; vmid < 16; vmid++) {
323 		bool valid;
324 
325 		valid = gmc_v11_0_get_vmid_pasid_mapping_info(adev, vmid,
326 							      &queried);
327 		if (!valid || queried != pasid)
328 			continue;
329 
330 		if (all_hub) {
331 			for_each_set_bit(i, adev->vmhubs_mask,
332 					 AMDGPU_MAX_VMHUBS)
333 				gmc_v11_0_flush_gpu_tlb(adev, vmid, i,
334 							flush_type);
335 		} else {
336 			gmc_v11_0_flush_gpu_tlb(adev, vmid, AMDGPU_GFXHUB(0),
337 						flush_type);
338 		}
339 	}
340 }
341 
342 static uint64_t gmc_v11_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
343 					     unsigned int vmid, uint64_t pd_addr)
344 {
345 	bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(ring->adev, ring->vm_hub);
346 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
347 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
348 	unsigned int eng = ring->vm_inv_eng;
349 
350 	/*
351 	 * It may lose gpuvm invalidate acknowldege state across power-gating
352 	 * off cycle, add semaphore acquire before invalidation and semaphore
353 	 * release after invalidation to avoid entering power gated state
354 	 * to WA the Issue
355 	 */
356 
357 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
358 	if (use_semaphore)
359 		/* a read return value of 1 means semaphore acuqire */
360 		amdgpu_ring_emit_reg_wait(ring,
361 					  hub->vm_inv_eng0_sem +
362 					  hub->eng_distance * eng, 0x1, 0x1);
363 
364 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
365 			      (hub->ctx_addr_distance * vmid),
366 			      lower_32_bits(pd_addr));
367 
368 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
369 			      (hub->ctx_addr_distance * vmid),
370 			      upper_32_bits(pd_addr));
371 
372 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
373 					    hub->eng_distance * eng,
374 					    hub->vm_inv_eng0_ack +
375 					    hub->eng_distance * eng,
376 					    req, 1 << vmid);
377 
378 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
379 	if (use_semaphore)
380 		/*
381 		 * add semaphore release after invalidation,
382 		 * write with 0 means semaphore release
383 		 */
384 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
385 				      hub->eng_distance * eng, 0);
386 
387 	return pd_addr;
388 }
389 
390 static void gmc_v11_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned int vmid,
391 					 unsigned int pasid)
392 {
393 	struct amdgpu_device *adev = ring->adev;
394 	uint32_t reg;
395 
396 	/* MES fw manages IH_VMID_x_LUT updating */
397 	if (ring->is_mes_queue)
398 		return;
399 
400 	if (ring->vm_hub == AMDGPU_GFXHUB(0))
401 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid;
402 	else
403 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT_MM) + vmid;
404 
405 	amdgpu_ring_emit_wreg(ring, reg, pasid);
406 }
407 
408 /*
409  * PTE format:
410  * 63:59 reserved
411  * 58:57 reserved
412  * 56 F
413  * 55 L
414  * 54 reserved
415  * 53:52 SW
416  * 51 T
417  * 50:48 mtype
418  * 47:12 4k physical page base address
419  * 11:7 fragment
420  * 6 write
421  * 5 read
422  * 4 exe
423  * 3 Z
424  * 2 snooped
425  * 1 system
426  * 0 valid
427  *
428  * PDE format:
429  * 63:59 block fragment size
430  * 58:55 reserved
431  * 54 P
432  * 53:48 reserved
433  * 47:6 physical base address of PD or PTE
434  * 5:3 reserved
435  * 2 C
436  * 1 system
437  * 0 valid
438  */
439 
440 static uint64_t gmc_v11_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
441 {
442 	switch (flags) {
443 	case AMDGPU_VM_MTYPE_DEFAULT:
444 		return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_NC);
445 	case AMDGPU_VM_MTYPE_NC:
446 		return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_NC);
447 	case AMDGPU_VM_MTYPE_WC:
448 		return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_WC);
449 	case AMDGPU_VM_MTYPE_CC:
450 		return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_CC);
451 	case AMDGPU_VM_MTYPE_UC:
452 		return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_UC);
453 	default:
454 		return AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_NC);
455 	}
456 }
457 
458 static void gmc_v11_0_get_vm_pde(struct amdgpu_device *adev, int level,
459 				 uint64_t *addr, uint64_t *flags)
460 {
461 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
462 		*addr = amdgpu_gmc_vram_mc2pa(adev, *addr);
463 	BUG_ON(*addr & 0xFFFF00000000003FULL);
464 
465 	if (!adev->gmc.translate_further)
466 		return;
467 
468 	if (level == AMDGPU_VM_PDB1) {
469 		/* Set the block fragment size */
470 		if (!(*flags & AMDGPU_PDE_PTE))
471 			*flags |= AMDGPU_PDE_BFS(0x9);
472 
473 	} else if (level == AMDGPU_VM_PDB0) {
474 		if (*flags & AMDGPU_PDE_PTE)
475 			*flags &= ~AMDGPU_PDE_PTE;
476 		else
477 			*flags |= AMDGPU_PTE_TF;
478 	}
479 }
480 
481 static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
482 				 struct amdgpu_bo_va_mapping *mapping,
483 				 uint64_t *flags)
484 {
485 	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
486 
487 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
488 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
489 
490 	*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
491 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
492 
493 	*flags &= ~AMDGPU_PTE_NOALLOC;
494 	*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
495 
496 	if (mapping->flags & AMDGPU_PTE_PRT) {
497 		*flags |= AMDGPU_PTE_PRT;
498 		*flags |= AMDGPU_PTE_SNOOPED;
499 		*flags |= AMDGPU_PTE_LOG;
500 		*flags |= AMDGPU_PTE_SYSTEM;
501 		*flags &= ~AMDGPU_PTE_VALID;
502 	}
503 
504 	if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
505 			       AMDGPU_GEM_CREATE_EXT_COHERENT |
506 			       AMDGPU_GEM_CREATE_UNCACHED))
507 		*flags = AMDGPU_PTE_MTYPE_NV10(*flags, MTYPE_UC);
508 }
509 
510 static unsigned int gmc_v11_0_get_vbios_fb_size(struct amdgpu_device *adev)
511 {
512 	u32 d1vga_control = RREG32_SOC15(DCE, 0, regD1VGA_CONTROL);
513 	unsigned int size;
514 
515 	if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
516 		size = AMDGPU_VBIOS_VGA_ALLOCATION;
517 	} else {
518 		u32 viewport;
519 		u32 pitch;
520 
521 		viewport = RREG32_SOC15(DCE, 0, regHUBP0_DCSURF_PRI_VIEWPORT_DIMENSION);
522 		pitch = RREG32_SOC15(DCE, 0, regHUBPREQ0_DCSURF_SURFACE_PITCH);
523 		size = (REG_GET_FIELD(viewport,
524 					HUBP0_DCSURF_PRI_VIEWPORT_DIMENSION, PRI_VIEWPORT_HEIGHT) *
525 				REG_GET_FIELD(pitch, HUBPREQ0_DCSURF_SURFACE_PITCH, PITCH) *
526 				4);
527 	}
528 
529 	return size;
530 }
531 
532 static const struct amdgpu_gmc_funcs gmc_v11_0_gmc_funcs = {
533 	.flush_gpu_tlb = gmc_v11_0_flush_gpu_tlb,
534 	.flush_gpu_tlb_pasid = gmc_v11_0_flush_gpu_tlb_pasid,
535 	.emit_flush_gpu_tlb = gmc_v11_0_emit_flush_gpu_tlb,
536 	.emit_pasid_mapping = gmc_v11_0_emit_pasid_mapping,
537 	.map_mtype = gmc_v11_0_map_mtype,
538 	.get_vm_pde = gmc_v11_0_get_vm_pde,
539 	.get_vm_pte = gmc_v11_0_get_vm_pte,
540 	.get_vbios_fb_size = gmc_v11_0_get_vbios_fb_size,
541 };
542 
543 static void gmc_v11_0_set_gmc_funcs(struct amdgpu_device *adev)
544 {
545 	adev->gmc.gmc_funcs = &gmc_v11_0_gmc_funcs;
546 }
547 
548 static void gmc_v11_0_set_umc_funcs(struct amdgpu_device *adev)
549 {
550 	switch (amdgpu_ip_version(adev, UMC_HWIP, 0)) {
551 	case IP_VERSION(8, 10, 0):
552 		adev->umc.channel_inst_num = UMC_V8_10_CHANNEL_INSTANCE_NUM;
553 		adev->umc.umc_inst_num = UMC_V8_10_UMC_INSTANCE_NUM;
554 		adev->umc.max_ras_err_cnt_per_query = UMC_V8_10_TOTAL_CHANNEL_NUM(adev);
555 		adev->umc.channel_offs = UMC_V8_10_PER_CHANNEL_OFFSET;
556 		adev->umc.retire_unit = UMC_V8_10_NA_COL_2BITS_POWER_OF_2_NUM;
557 		if (adev->umc.node_inst_num == 4)
558 			adev->umc.channel_idx_tbl = &umc_v8_10_channel_idx_tbl_ext0[0][0][0];
559 		else
560 			adev->umc.channel_idx_tbl = &umc_v8_10_channel_idx_tbl[0][0][0];
561 		adev->umc.ras = &umc_v8_10_ras;
562 		break;
563 	case IP_VERSION(8, 11, 0):
564 		break;
565 	default:
566 		break;
567 	}
568 }
569 
570 
571 static void gmc_v11_0_set_mmhub_funcs(struct amdgpu_device *adev)
572 {
573 	switch (amdgpu_ip_version(adev, MMHUB_HWIP, 0)) {
574 	case IP_VERSION(3, 0, 1):
575 		adev->mmhub.funcs = &mmhub_v3_0_1_funcs;
576 		break;
577 	case IP_VERSION(3, 0, 2):
578 		adev->mmhub.funcs = &mmhub_v3_0_2_funcs;
579 		break;
580 	case IP_VERSION(3, 3, 0):
581 	case IP_VERSION(3, 3, 1):
582 		adev->mmhub.funcs = &mmhub_v3_3_funcs;
583 		break;
584 	default:
585 		adev->mmhub.funcs = &mmhub_v3_0_funcs;
586 		break;
587 	}
588 }
589 
590 static void gmc_v11_0_set_gfxhub_funcs(struct amdgpu_device *adev)
591 {
592 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
593 	case IP_VERSION(11, 0, 3):
594 		adev->gfxhub.funcs = &gfxhub_v3_0_3_funcs;
595 		break;
596 	case IP_VERSION(11, 5, 0):
597 	case IP_VERSION(11, 5, 1):
598 	case IP_VERSION(11, 5, 2):
599 		adev->gfxhub.funcs = &gfxhub_v11_5_0_funcs;
600 		break;
601 	default:
602 		adev->gfxhub.funcs = &gfxhub_v3_0_funcs;
603 		break;
604 	}
605 }
606 
607 static int gmc_v11_0_early_init(struct amdgpu_ip_block *ip_block)
608 {
609 	struct amdgpu_device *adev = ip_block->adev;
610 
611 	gmc_v11_0_set_gfxhub_funcs(adev);
612 	gmc_v11_0_set_mmhub_funcs(adev);
613 	gmc_v11_0_set_gmc_funcs(adev);
614 	gmc_v11_0_set_irq_funcs(adev);
615 	gmc_v11_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 	adev->gmc.noretry_flags = AMDGPU_VM_NORETRY_FLAGS_TF;
624 
625 	return 0;
626 }
627 
628 static int gmc_v11_0_late_init(struct amdgpu_ip_block *ip_block)
629 {
630 	struct amdgpu_device *adev = ip_block->adev;
631 	int r;
632 
633 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
634 	if (r)
635 		return r;
636 
637 	r = amdgpu_gmc_ras_late_init(adev);
638 	if (r)
639 		return r;
640 
641 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
642 }
643 
644 static void gmc_v11_0_vram_gtt_location(struct amdgpu_device *adev,
645 					struct amdgpu_gmc *mc)
646 {
647 	u64 base = 0;
648 
649 	base = adev->mmhub.funcs->get_fb_location(adev);
650 
651 	amdgpu_gmc_set_agp_default(adev, mc);
652 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
653 	amdgpu_gmc_gart_location(adev, mc, AMDGPU_GART_PLACEMENT_HIGH);
654 	if (!amdgpu_sriov_vf(adev) &&
655 	    (amdgpu_ip_version(adev, GC_HWIP, 0) < IP_VERSION(11, 5, 0)) &&
656 	    (amdgpu_agp == 1))
657 		amdgpu_gmc_agp_location(adev, mc);
658 
659 	/* base offset of vram pages */
660 	if (amdgpu_sriov_vf(adev))
661 		adev->vm_manager.vram_base_offset = 0;
662 	else
663 		adev->vm_manager.vram_base_offset = adev->mmhub.funcs->get_mc_fb_offset(adev);
664 }
665 
666 /**
667  * gmc_v11_0_mc_init - initialize the memory controller driver params
668  *
669  * @adev: amdgpu_device pointer
670  *
671  * Look up the amount of vram, vram width, and decide how to place
672  * vram and gart within the GPU's physical address space.
673  * Returns 0 for success.
674  */
675 static int gmc_v11_0_mc_init(struct amdgpu_device *adev)
676 {
677 	int r;
678 
679 	/* size in MB on si */
680 	adev->gmc.mc_vram_size =
681 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
682 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
683 
684 	if (!(adev->flags & AMD_IS_APU)) {
685 		r = amdgpu_device_resize_fb_bar(adev);
686 		if (r)
687 			return r;
688 	}
689 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
690 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
691 
692 #ifdef CONFIG_X86_64
693 	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) {
694 		adev->gmc.aper_base = adev->mmhub.funcs->get_mc_fb_offset(adev);
695 		adev->gmc.aper_size = adev->gmc.real_vram_size;
696 	}
697 #endif
698 	/* In case the PCI BAR is larger than the actual amount of vram */
699 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
700 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
701 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
702 
703 	/* set the gart size */
704 	if (amdgpu_gart_size == -1)
705 		adev->gmc.gart_size = 512ULL << 20;
706 	else
707 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
708 
709 	gmc_v11_0_vram_gtt_location(adev, &adev->gmc);
710 
711 	return 0;
712 }
713 
714 static int gmc_v11_0_gart_init(struct amdgpu_device *adev)
715 {
716 	int r;
717 
718 	if (adev->gart.bo) {
719 		WARN(1, "PCIE GART already initialized\n");
720 		return 0;
721 	}
722 
723 	/* Initialize common gart structure */
724 	r = amdgpu_gart_init(adev);
725 	if (r)
726 		return r;
727 
728 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
729 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(0ULL, MTYPE_UC) |
730 				 AMDGPU_PTE_EXECUTABLE;
731 
732 	return amdgpu_gart_table_vram_alloc(adev);
733 }
734 
735 static int gmc_v11_0_sw_init(struct amdgpu_ip_block *ip_block)
736 {
737 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
738 	struct amdgpu_device *adev = ip_block->adev;
739 
740 	adev->mmhub.funcs->init(adev);
741 
742 	adev->gfxhub.funcs->init(adev);
743 
744 	spin_lock_init(&adev->gmc.invalidate_lock);
745 
746 	r = amdgpu_atomfirmware_get_vram_info(adev,
747 					      &vram_width, &vram_type, &vram_vendor);
748 	adev->gmc.vram_width = vram_width;
749 
750 	adev->gmc.vram_type = vram_type;
751 	adev->gmc.vram_vendor = vram_vendor;
752 
753 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
754 	case IP_VERSION(11, 0, 0):
755 	case IP_VERSION(11, 0, 1):
756 	case IP_VERSION(11, 0, 2):
757 	case IP_VERSION(11, 0, 3):
758 	case IP_VERSION(11, 0, 4):
759 	case IP_VERSION(11, 5, 0):
760 	case IP_VERSION(11, 5, 1):
761 	case IP_VERSION(11, 5, 2):
762 		set_bit(AMDGPU_GFXHUB(0), adev->vmhubs_mask);
763 		set_bit(AMDGPU_MMHUB0(0), adev->vmhubs_mask);
764 		/*
765 		 * To fulfill 4-level page support,
766 		 * vm size is 256TB (48bit), maximum size,
767 		 * block size 512 (9bit)
768 		 */
769 		amdgpu_vm_adjust_size(adev, 256 * 1024, 9, 3, 48);
770 		break;
771 	default:
772 		break;
773 	}
774 
775 	/* This interrupt is VMC page fault.*/
776 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_VMC,
777 			      VMC_1_0__SRCID__VM_FAULT,
778 			      &adev->gmc.vm_fault);
779 
780 	if (r)
781 		return r;
782 
783 	r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX,
784 			      UTCL2_1_0__SRCID__FAULT,
785 			      &adev->gmc.vm_fault);
786 	if (r)
787 		return r;
788 
789 	if (!amdgpu_sriov_vf(adev)) {
790 		/* interrupt sent to DF. */
791 		r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_DF, 0,
792 				      &adev->gmc.ecc_irq);
793 		if (r)
794 			return r;
795 	}
796 
797 	/*
798 	 * Set the internal MC address mask This is the max address of the GPU's
799 	 * internal address space.
800 	 */
801 	adev->gmc.mc_mask = 0xffffffffffffULL; /* 48 bit MC */
802 
803 	r = dma_set_mask_and_coherent(adev->dev, DMA_BIT_MASK(44));
804 	if (r) {
805 		dev_warn(adev->dev, "amdgpu: No suitable DMA available.\n");
806 		return r;
807 	}
808 
809 	adev->need_swiotlb = drm_need_swiotlb(44);
810 
811 	r = gmc_v11_0_mc_init(adev);
812 	if (r)
813 		return r;
814 
815 	amdgpu_gmc_get_vbios_allocations(adev);
816 
817 	/* Memory manager */
818 	r = amdgpu_bo_init(adev);
819 	if (r)
820 		return r;
821 
822 	r = gmc_v11_0_gart_init(adev);
823 	if (r)
824 		return r;
825 
826 	/*
827 	 * number of VMs
828 	 * VMID 0 is reserved for System
829 	 * amdgpu graphics/compute will use VMIDs 1-7
830 	 * amdkfd will use VMIDs 8-15
831 	 */
832 	adev->vm_manager.first_kfd_vmid = 8;
833 
834 	amdgpu_vm_manager_init(adev);
835 
836 	r = amdgpu_gmc_ras_sw_init(adev);
837 	if (r)
838 		return r;
839 
840 	return 0;
841 }
842 
843 /**
844  * gmc_v11_0_gart_fini - vm fini callback
845  *
846  * @adev: amdgpu_device pointer
847  *
848  * Tears down the driver GART/VM setup (CIK).
849  */
850 static void gmc_v11_0_gart_fini(struct amdgpu_device *adev)
851 {
852 	amdgpu_gart_table_vram_free(adev);
853 }
854 
855 static int gmc_v11_0_sw_fini(struct amdgpu_ip_block *ip_block)
856 {
857 	struct amdgpu_device *adev = ip_block->adev;
858 
859 	amdgpu_vm_manager_fini(adev);
860 	gmc_v11_0_gart_fini(adev);
861 	amdgpu_gem_force_release(adev);
862 	amdgpu_bo_fini(adev);
863 
864 	return 0;
865 }
866 
867 static void gmc_v11_0_init_golden_registers(struct amdgpu_device *adev)
868 {
869 	if (amdgpu_sriov_vf(adev)) {
870 		struct amdgpu_vmhub *hub = &adev->vmhub[AMDGPU_MMHUB0(0)];
871 
872 		WREG32(hub->vm_contexts_disable, 0);
873 		return;
874 	}
875 }
876 
877 /**
878  * gmc_v11_0_gart_enable - gart enable
879  *
880  * @adev: amdgpu_device pointer
881  */
882 static int gmc_v11_0_gart_enable(struct amdgpu_device *adev)
883 {
884 	int r;
885 	bool value;
886 
887 	if (adev->gart.bo == NULL) {
888 		dev_err(adev->dev, "No VRAM object for PCIE GART.\n");
889 		return -EINVAL;
890 	}
891 
892 	amdgpu_gtt_mgr_recover(&adev->mman.gtt_mgr);
893 
894 	r = adev->mmhub.funcs->gart_enable(adev);
895 	if (r)
896 		return r;
897 
898 	/* Flush HDP after it is initialized */
899 	adev->hdp.funcs->flush_hdp(adev, NULL);
900 
901 	value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ?
902 		false : true;
903 
904 	adev->mmhub.funcs->set_fault_enable_default(adev, value);
905 	gmc_v11_0_flush_gpu_tlb(adev, 0, AMDGPU_MMHUB0(0), 0);
906 
907 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
908 		 (unsigned int)(adev->gmc.gart_size >> 20),
909 		 (unsigned long long)amdgpu_bo_gpu_offset(adev->gart.bo));
910 
911 	return 0;
912 }
913 
914 static int gmc_v11_0_hw_init(struct amdgpu_ip_block *ip_block)
915 {
916 	struct amdgpu_device *adev = ip_block->adev;
917 	int r;
918 
919 	adev->gmc.flush_pasid_uses_kiq = !amdgpu_emu_mode;
920 
921 	/* The sequence of these two function calls matters.*/
922 	gmc_v11_0_init_golden_registers(adev);
923 
924 	r = gmc_v11_0_gart_enable(adev);
925 	if (r)
926 		return r;
927 
928 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
929 		adev->umc.funcs->init_registers(adev);
930 
931 	return 0;
932 }
933 
934 /**
935  * gmc_v11_0_gart_disable - gart disable
936  *
937  * @adev: amdgpu_device pointer
938  *
939  * This disables all VM page table.
940  */
941 static void gmc_v11_0_gart_disable(struct amdgpu_device *adev)
942 {
943 	adev->mmhub.funcs->gart_disable(adev);
944 }
945 
946 static int gmc_v11_0_hw_fini(struct amdgpu_ip_block *ip_block)
947 {
948 	struct amdgpu_device *adev = ip_block->adev;
949 
950 	if (amdgpu_sriov_vf(adev)) {
951 		/* full access mode, so don't touch any GMC register */
952 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
953 		return 0;
954 	}
955 
956 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
957 
958 	if (adev->gmc.ecc_irq.funcs &&
959 		amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__UMC))
960 		amdgpu_irq_put(adev, &adev->gmc.ecc_irq, 0);
961 
962 	gmc_v11_0_gart_disable(adev);
963 
964 	return 0;
965 }
966 
967 static int gmc_v11_0_suspend(struct amdgpu_ip_block *ip_block)
968 {
969 	gmc_v11_0_hw_fini(ip_block);
970 
971 	return 0;
972 }
973 
974 static int gmc_v11_0_resume(struct amdgpu_ip_block *ip_block)
975 {
976 	int r;
977 
978 	r = gmc_v11_0_hw_init(ip_block);
979 	if (r)
980 		return r;
981 
982 	amdgpu_vmid_reset_all(ip_block->adev);
983 
984 	return 0;
985 }
986 
987 static bool gmc_v11_0_is_idle(void *handle)
988 {
989 	/* MC is always ready in GMC v11.*/
990 	return true;
991 }
992 
993 static int gmc_v11_0_wait_for_idle(struct amdgpu_ip_block *ip_block)
994 {
995 	/* There is no need to wait for MC idle in GMC v11.*/
996 	return 0;
997 }
998 
999 static int gmc_v11_0_set_clockgating_state(void *handle,
1000 					   enum amd_clockgating_state state)
1001 {
1002 	int r;
1003 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1004 
1005 	r = adev->mmhub.funcs->set_clockgating(adev, state);
1006 	if (r)
1007 		return r;
1008 
1009 	return athub_v3_0_set_clockgating(adev, state);
1010 }
1011 
1012 static void gmc_v11_0_get_clockgating_state(void *handle, u64 *flags)
1013 {
1014 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1015 
1016 	adev->mmhub.funcs->get_clockgating(adev, flags);
1017 
1018 	athub_v3_0_get_clockgating(adev, flags);
1019 }
1020 
1021 static int gmc_v11_0_set_powergating_state(void *handle,
1022 					   enum amd_powergating_state state)
1023 {
1024 	return 0;
1025 }
1026 
1027 const struct amd_ip_funcs gmc_v11_0_ip_funcs = {
1028 	.name = "gmc_v11_0",
1029 	.early_init = gmc_v11_0_early_init,
1030 	.sw_init = gmc_v11_0_sw_init,
1031 	.hw_init = gmc_v11_0_hw_init,
1032 	.late_init = gmc_v11_0_late_init,
1033 	.sw_fini = gmc_v11_0_sw_fini,
1034 	.hw_fini = gmc_v11_0_hw_fini,
1035 	.suspend = gmc_v11_0_suspend,
1036 	.resume = gmc_v11_0_resume,
1037 	.is_idle = gmc_v11_0_is_idle,
1038 	.wait_for_idle = gmc_v11_0_wait_for_idle,
1039 	.set_clockgating_state = gmc_v11_0_set_clockgating_state,
1040 	.set_powergating_state = gmc_v11_0_set_powergating_state,
1041 	.get_clockgating_state = gmc_v11_0_get_clockgating_state,
1042 };
1043 
1044 const struct amdgpu_ip_block_version gmc_v11_0_ip_block = {
1045 	.type = AMD_IP_BLOCK_TYPE_GMC,
1046 	.major = 11,
1047 	.minor = 0,
1048 	.rev = 0,
1049 	.funcs = &gmc_v11_0_ip_funcs,
1050 };
1051