xref: /linux/drivers/gpu/drm/amd/amdgpu/gmc_v11_0.c (revision 2f804aca48322f02a8f44cca540663845ee80fb1)
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 "oss/osssys_6_0_0_offset.h"
35 #include "ivsrcid/vmc/irqsrcs_vmc_1_0.h"
36 #include "navi10_enum.h"
37 #include "soc15.h"
38 #include "soc15d.h"
39 #include "soc15_common.h"
40 #include "nbio_v4_3.h"
41 #include "gfxhub_v3_0.h"
42 #include "gfxhub_v3_0_3.h"
43 #include "mmhub_v3_0.h"
44 #include "mmhub_v3_0_1.h"
45 #include "mmhub_v3_0_2.h"
46 #include "athub_v3_0.h"
47 
48 
49 static int gmc_v11_0_ecc_interrupt_state(struct amdgpu_device *adev,
50 					 struct amdgpu_irq_src *src,
51 					 unsigned type,
52 					 enum amdgpu_interrupt_state state)
53 {
54 	return 0;
55 }
56 
57 static int
58 gmc_v11_0_vm_fault_interrupt_state(struct amdgpu_device *adev,
59 				   struct amdgpu_irq_src *src, unsigned type,
60 				   enum amdgpu_interrupt_state state)
61 {
62 	switch (state) {
63 	case AMDGPU_IRQ_STATE_DISABLE:
64 		/* MM HUB */
65 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, false);
66 		/* GFX HUB */
67 		/* This works because this interrupt is only
68 		 * enabled at init/resume and disabled in
69 		 * fini/suspend, so the overall state doesn't
70 		 * change over the course of suspend/resume.
71 		 */
72 		if (!adev->in_s0ix)
73 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, false);
74 		break;
75 	case AMDGPU_IRQ_STATE_ENABLE:
76 		/* MM HUB */
77 		amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_MMHUB_0, true);
78 		/* GFX HUB */
79 		/* This works because this interrupt is only
80 		 * enabled at init/resume and disabled in
81 		 * fini/suspend, so the overall state doesn't
82 		 * change over the course of suspend/resume.
83 		 */
84 		if (!adev->in_s0ix)
85 			amdgpu_gmc_set_vm_fault_masks(adev, AMDGPU_GFXHUB_0, true);
86 		break;
87 	default:
88 		break;
89 	}
90 
91 	return 0;
92 }
93 
94 static int gmc_v11_0_process_interrupt(struct amdgpu_device *adev,
95 				       struct amdgpu_irq_src *source,
96 				       struct amdgpu_iv_entry *entry)
97 {
98 	struct amdgpu_vmhub *hub = &adev->vmhub[entry->vmid_src];
99 	uint32_t status = 0;
100 	u64 addr;
101 
102 	addr = (u64)entry->src_data[0] << 12;
103 	addr |= ((u64)entry->src_data[1] & 0xf) << 44;
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 
118 	if (printk_ratelimit()) {
119 		struct amdgpu_task_info task_info;
120 
121 		memset(&task_info, 0, sizeof(struct amdgpu_task_info));
122 		amdgpu_vm_get_task_info(adev, entry->pasid, &task_info);
123 
124 		dev_err(adev->dev,
125 			"[%s] page fault (src_id:%u ring:%u vmid:%u pasid:%u, "
126 			"for process %s pid %d thread %s pid %d)\n",
127 			entry->vmid_src ? "mmhub" : "gfxhub",
128 			entry->src_id, entry->ring_id, entry->vmid,
129 			entry->pasid, task_info.process_name, task_info.tgid,
130 			task_info.task_name, task_info.pid);
131 		dev_err(adev->dev, "  in page starting at address 0x%016llx from client %d\n",
132 			addr, entry->client_id);
133 		if (!amdgpu_sriov_vf(adev))
134 			hub->vmhub_funcs->print_l2_protection_fault_status(adev, status);
135 	}
136 
137 	return 0;
138 }
139 
140 static const struct amdgpu_irq_src_funcs gmc_v11_0_irq_funcs = {
141 	.set = gmc_v11_0_vm_fault_interrupt_state,
142 	.process = gmc_v11_0_process_interrupt,
143 };
144 
145 static const struct amdgpu_irq_src_funcs gmc_v11_0_ecc_funcs = {
146 	.set = gmc_v11_0_ecc_interrupt_state,
147 	.process = amdgpu_umc_process_ecc_irq,
148 };
149 
150 static void gmc_v11_0_set_irq_funcs(struct amdgpu_device *adev)
151 {
152 	adev->gmc.vm_fault.num_types = 1;
153 	adev->gmc.vm_fault.funcs = &gmc_v11_0_irq_funcs;
154 
155 	if (!amdgpu_sriov_vf(adev)) {
156 		adev->gmc.ecc_irq.num_types = 1;
157 		adev->gmc.ecc_irq.funcs = &gmc_v11_0_ecc_funcs;
158 	}
159 }
160 
161 /**
162  * gmc_v11_0_use_invalidate_semaphore - judge whether to use semaphore
163  *
164  * @adev: amdgpu_device pointer
165  * @vmhub: vmhub type
166  *
167  */
168 static bool gmc_v11_0_use_invalidate_semaphore(struct amdgpu_device *adev,
169 				       uint32_t vmhub)
170 {
171 	return ((vmhub == AMDGPU_MMHUB_0) &&
172 		(!amdgpu_sriov_vf(adev)));
173 }
174 
175 static bool gmc_v11_0_get_vmid_pasid_mapping_info(
176 					struct amdgpu_device *adev,
177 					uint8_t vmid, uint16_t *p_pasid)
178 {
179 	*p_pasid = RREG32(SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid) & 0xffff;
180 
181 	return !!(*p_pasid);
182 }
183 
184 /*
185  * GART
186  * VMID 0 is the physical GPU addresses as used by the kernel.
187  * VMIDs 1-15 are used for userspace clients and are handled
188  * by the amdgpu vm/hsa code.
189  */
190 
191 static void gmc_v11_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
192 				   unsigned int vmhub, uint32_t flush_type)
193 {
194 	bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(adev, vmhub);
195 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
196 	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
197 	u32 tmp;
198 	/* Use register 17 for GART */
199 	const unsigned eng = 17;
200 	unsigned int i;
201 	unsigned char hub_ip = 0;
202 
203 	hub_ip = (vmhub == AMDGPU_GFXHUB_0) ?
204 		   GC_HWIP : MMHUB_HWIP;
205 
206 	spin_lock(&adev->gmc.invalidate_lock);
207 	/*
208 	 * It may lose gpuvm invalidate acknowldege state across power-gating
209 	 * off cycle, add semaphore acquire before invalidation and semaphore
210 	 * release after invalidation to avoid entering power gated state
211 	 * to WA the Issue
212 	 */
213 
214 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
215 	if (use_semaphore) {
216 		for (i = 0; i < adev->usec_timeout; i++) {
217 			/* a read return value of 1 means semaphore acuqire */
218 			tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
219 					    hub->eng_distance * eng, hub_ip);
220 			if (tmp & 0x1)
221 				break;
222 			udelay(1);
223 		}
224 
225 		if (i >= adev->usec_timeout)
226 			DRM_ERROR("Timeout waiting for sem acquire in VM flush!\n");
227 	}
228 
229 	WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_req + hub->eng_distance * eng, inv_req, hub_ip);
230 
231 	/* Wait for ACK with a delay.*/
232 	for (i = 0; i < adev->usec_timeout; i++) {
233 		tmp = RREG32_RLC_NO_KIQ(hub->vm_inv_eng0_ack +
234 				    hub->eng_distance * eng, hub_ip);
235 		tmp &= 1 << vmid;
236 		if (tmp)
237 			break;
238 
239 		udelay(1);
240 	}
241 
242 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
243 	if (use_semaphore)
244 		/*
245 		 * add semaphore release after invalidation,
246 		 * write with 0 means semaphore release
247 		 */
248 		WREG32_RLC_NO_KIQ(hub->vm_inv_eng0_sem +
249 			      hub->eng_distance * eng, 0, hub_ip);
250 
251 	/* Issue additional private vm invalidation to MMHUB */
252 	if ((vmhub != AMDGPU_GFXHUB_0) &&
253 	    (hub->vm_l2_bank_select_reserved_cid2) &&
254 		!amdgpu_sriov_vf(adev)) {
255 		inv_req = RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
256 		/* bit 25: RSERVED_CACHE_PRIVATE_INVALIDATION */
257 		inv_req |= (1 << 25);
258 		/* Issue private invalidation */
259 		WREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2, inv_req);
260 		/* Read back to ensure invalidation is done*/
261 		RREG32_NO_KIQ(hub->vm_l2_bank_select_reserved_cid2);
262 	}
263 
264 	spin_unlock(&adev->gmc.invalidate_lock);
265 
266 	if (i < adev->usec_timeout)
267 		return;
268 
269 	DRM_ERROR("Timeout waiting for VM flush ACK!\n");
270 }
271 
272 /**
273  * gmc_v11_0_flush_gpu_tlb - gart tlb flush callback
274  *
275  * @adev: amdgpu_device pointer
276  * @vmid: vm instance to flush
277  * @vmhub: which hub to flush
278  * @flush_type: the flush type
279  *
280  * Flush the TLB for the requested page table.
281  */
282 static void gmc_v11_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
283 					uint32_t vmhub, uint32_t flush_type)
284 {
285 	if ((vmhub == AMDGPU_GFXHUB_0) && !adev->gfx.is_poweron)
286 		return;
287 
288 	/* flush hdp cache */
289 	adev->hdp.funcs->flush_hdp(adev, NULL);
290 
291 	/* For SRIOV run time, driver shouldn't access the register through MMIO
292 	 * Directly use kiq to do the vm invalidation instead
293 	 */
294 	if ((adev->gfx.kiq.ring.sched.ready || adev->mes.ring.sched.ready) &&
295 	    (amdgpu_sriov_runtime(adev) || !amdgpu_sriov_vf(adev))) {
296 		struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
297 		const unsigned eng = 17;
298 		u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
299 		u32 req = hub->vm_inv_eng0_req + hub->eng_distance * eng;
300 		u32 ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;
301 
302 		amdgpu_virt_kiq_reg_write_reg_wait(adev, req, ack, inv_req,
303 				1 << vmid);
304 		return;
305 	}
306 
307 	mutex_lock(&adev->mman.gtt_window_lock);
308 	gmc_v11_0_flush_vm_hub(adev, vmid, vmhub, 0);
309 	mutex_unlock(&adev->mman.gtt_window_lock);
310 	return;
311 }
312 
313 /**
314  * gmc_v11_0_flush_gpu_tlb_pasid - tlb flush via pasid
315  *
316  * @adev: amdgpu_device pointer
317  * @pasid: pasid to be flush
318  * @flush_type: the flush type
319  * @all_hub: flush all hubs
320  *
321  * Flush the TLB for the requested pasid.
322  */
323 static int gmc_v11_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
324 					uint16_t pasid, uint32_t flush_type,
325 					bool all_hub)
326 {
327 	int vmid, i;
328 	signed long r;
329 	uint32_t seq;
330 	uint16_t queried_pasid;
331 	bool ret;
332 	struct amdgpu_ring *ring = &adev->gfx.kiq.ring;
333 	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
334 
335 	if (amdgpu_emu_mode == 0 && ring->sched.ready) {
336 		spin_lock(&adev->gfx.kiq.ring_lock);
337 		/* 2 dwords flush + 8 dwords fence */
338 		amdgpu_ring_alloc(ring, kiq->pmf->invalidate_tlbs_size + 8);
339 		kiq->pmf->kiq_invalidate_tlbs(ring,
340 					pasid, flush_type, all_hub);
341 		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
342 		if (r) {
343 			amdgpu_ring_undo(ring);
344 			spin_unlock(&adev->gfx.kiq.ring_lock);
345 			return -ETIME;
346 		}
347 
348 		amdgpu_ring_commit(ring);
349 		spin_unlock(&adev->gfx.kiq.ring_lock);
350 		r = amdgpu_fence_wait_polling(ring, seq, adev->usec_timeout);
351 		if (r < 1) {
352 			dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
353 			return -ETIME;
354 		}
355 
356 		return 0;
357 	}
358 
359 	for (vmid = 1; vmid < 16; vmid++) {
360 
361 		ret = gmc_v11_0_get_vmid_pasid_mapping_info(adev, vmid,
362 				&queried_pasid);
363 		if (ret	&& queried_pasid == pasid) {
364 			if (all_hub) {
365 				for (i = 0; i < adev->num_vmhubs; i++)
366 					gmc_v11_0_flush_gpu_tlb(adev, vmid,
367 							i, flush_type);
368 			} else {
369 				gmc_v11_0_flush_gpu_tlb(adev, vmid,
370 						AMDGPU_GFXHUB_0, flush_type);
371 			}
372 		}
373 	}
374 
375 	return 0;
376 }
377 
378 static uint64_t gmc_v11_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
379 					     unsigned vmid, uint64_t pd_addr)
380 {
381 	bool use_semaphore = gmc_v11_0_use_invalidate_semaphore(ring->adev, ring->vm_hub);
382 	struct amdgpu_vmhub *hub = &ring->adev->vmhub[ring->vm_hub];
383 	uint32_t req = hub->vmhub_funcs->get_invalidate_req(vmid, 0);
384 	unsigned eng = ring->vm_inv_eng;
385 
386 	/*
387 	 * It may lose gpuvm invalidate acknowldege state across power-gating
388 	 * off cycle, add semaphore acquire before invalidation and semaphore
389 	 * release after invalidation to avoid entering power gated state
390 	 * to WA the Issue
391 	 */
392 
393 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
394 	if (use_semaphore)
395 		/* a read return value of 1 means semaphore acuqire */
396 		amdgpu_ring_emit_reg_wait(ring,
397 					  hub->vm_inv_eng0_sem +
398 					  hub->eng_distance * eng, 0x1, 0x1);
399 
400 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_lo32 +
401 			      (hub->ctx_addr_distance * vmid),
402 			      lower_32_bits(pd_addr));
403 
404 	amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 +
405 			      (hub->ctx_addr_distance * vmid),
406 			      upper_32_bits(pd_addr));
407 
408 	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
409 					    hub->eng_distance * eng,
410 					    hub->vm_inv_eng0_ack +
411 					    hub->eng_distance * eng,
412 					    req, 1 << vmid);
413 
414 	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
415 	if (use_semaphore)
416 		/*
417 		 * add semaphore release after invalidation,
418 		 * write with 0 means semaphore release
419 		 */
420 		amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_sem +
421 				      hub->eng_distance * eng, 0);
422 
423 	return pd_addr;
424 }
425 
426 static void gmc_v11_0_emit_pasid_mapping(struct amdgpu_ring *ring, unsigned vmid,
427 					 unsigned pasid)
428 {
429 	struct amdgpu_device *adev = ring->adev;
430 	uint32_t reg;
431 
432 	/* MES fw manages IH_VMID_x_LUT updating */
433 	if (ring->is_mes_queue)
434 		return;
435 
436 	if (ring->vm_hub == AMDGPU_GFXHUB_0)
437 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT) + vmid;
438 	else
439 		reg = SOC15_REG_OFFSET(OSSSYS, 0, regIH_VMID_0_LUT_MM) + vmid;
440 
441 	amdgpu_ring_emit_wreg(ring, reg, pasid);
442 }
443 
444 /*
445  * PTE format:
446  * 63:59 reserved
447  * 58:57 reserved
448  * 56 F
449  * 55 L
450  * 54 reserved
451  * 53:52 SW
452  * 51 T
453  * 50:48 mtype
454  * 47:12 4k physical page base address
455  * 11:7 fragment
456  * 6 write
457  * 5 read
458  * 4 exe
459  * 3 Z
460  * 2 snooped
461  * 1 system
462  * 0 valid
463  *
464  * PDE format:
465  * 63:59 block fragment size
466  * 58:55 reserved
467  * 54 P
468  * 53:48 reserved
469  * 47:6 physical base address of PD or PTE
470  * 5:3 reserved
471  * 2 C
472  * 1 system
473  * 0 valid
474  */
475 
476 static uint64_t gmc_v11_0_map_mtype(struct amdgpu_device *adev, uint32_t flags)
477 {
478 	switch (flags) {
479 	case AMDGPU_VM_MTYPE_DEFAULT:
480 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
481 	case AMDGPU_VM_MTYPE_NC:
482 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
483 	case AMDGPU_VM_MTYPE_WC:
484 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_WC);
485 	case AMDGPU_VM_MTYPE_CC:
486 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_CC);
487 	case AMDGPU_VM_MTYPE_UC:
488 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
489 	default:
490 		return AMDGPU_PTE_MTYPE_NV10(MTYPE_NC);
491 	}
492 }
493 
494 static void gmc_v11_0_get_vm_pde(struct amdgpu_device *adev, int level,
495 				 uint64_t *addr, uint64_t *flags)
496 {
497 	if (!(*flags & AMDGPU_PDE_PTE) && !(*flags & AMDGPU_PTE_SYSTEM))
498 		*addr = adev->vm_manager.vram_base_offset + *addr -
499 			adev->gmc.vram_start;
500 	BUG_ON(*addr & 0xFFFF00000000003FULL);
501 
502 	if (!adev->gmc.translate_further)
503 		return;
504 
505 	if (level == AMDGPU_VM_PDB1) {
506 		/* Set the block fragment size */
507 		if (!(*flags & AMDGPU_PDE_PTE))
508 			*flags |= AMDGPU_PDE_BFS(0x9);
509 
510 	} else if (level == AMDGPU_VM_PDB0) {
511 		if (*flags & AMDGPU_PDE_PTE)
512 			*flags &= ~AMDGPU_PDE_PTE;
513 		else
514 			*flags |= AMDGPU_PTE_TF;
515 	}
516 }
517 
518 static void gmc_v11_0_get_vm_pte(struct amdgpu_device *adev,
519 				 struct amdgpu_bo_va_mapping *mapping,
520 				 uint64_t *flags)
521 {
522 	struct amdgpu_bo *bo = mapping->bo_va->base.bo;
523 
524 	*flags &= ~AMDGPU_PTE_EXECUTABLE;
525 	*flags |= mapping->flags & AMDGPU_PTE_EXECUTABLE;
526 
527 	*flags &= ~AMDGPU_PTE_MTYPE_NV10_MASK;
528 	*flags |= (mapping->flags & AMDGPU_PTE_MTYPE_NV10_MASK);
529 
530 	*flags &= ~AMDGPU_PTE_NOALLOC;
531 	*flags |= (mapping->flags & AMDGPU_PTE_NOALLOC);
532 
533 	if (mapping->flags & AMDGPU_PTE_PRT) {
534 		*flags |= AMDGPU_PTE_PRT;
535 		*flags |= AMDGPU_PTE_SNOOPED;
536 		*flags |= AMDGPU_PTE_LOG;
537 		*flags |= AMDGPU_PTE_SYSTEM;
538 		*flags &= ~AMDGPU_PTE_VALID;
539 	}
540 
541 	if (bo && bo->flags & (AMDGPU_GEM_CREATE_COHERENT |
542 			       AMDGPU_GEM_CREATE_UNCACHED))
543 		*flags = (*flags & ~AMDGPU_PTE_MTYPE_NV10_MASK) |
544 			 AMDGPU_PTE_MTYPE_NV10(MTYPE_UC);
545 }
546 
547 static unsigned gmc_v11_0_get_vbios_fb_size(struct amdgpu_device *adev)
548 {
549 	return 0;
550 }
551 
552 static const struct amdgpu_gmc_funcs gmc_v11_0_gmc_funcs = {
553 	.flush_gpu_tlb = gmc_v11_0_flush_gpu_tlb,
554 	.flush_gpu_tlb_pasid = gmc_v11_0_flush_gpu_tlb_pasid,
555 	.emit_flush_gpu_tlb = gmc_v11_0_emit_flush_gpu_tlb,
556 	.emit_pasid_mapping = gmc_v11_0_emit_pasid_mapping,
557 	.map_mtype = gmc_v11_0_map_mtype,
558 	.get_vm_pde = gmc_v11_0_get_vm_pde,
559 	.get_vm_pte = gmc_v11_0_get_vm_pte,
560 	.get_vbios_fb_size = gmc_v11_0_get_vbios_fb_size,
561 };
562 
563 static void gmc_v11_0_set_gmc_funcs(struct amdgpu_device *adev)
564 {
565 	adev->gmc.gmc_funcs = &gmc_v11_0_gmc_funcs;
566 }
567 
568 static void gmc_v11_0_set_umc_funcs(struct amdgpu_device *adev)
569 {
570 	switch (adev->ip_versions[UMC_HWIP][0]) {
571 	case IP_VERSION(8, 10, 0):
572 		adev->umc.channel_inst_num = UMC_V8_10_CHANNEL_INSTANCE_NUM;
573 		adev->umc.umc_inst_num = UMC_V8_10_UMC_INSTANCE_NUM;
574 		adev->umc.max_ras_err_cnt_per_query = UMC_V8_10_TOTAL_CHANNEL_NUM(adev);
575 		adev->umc.channel_offs = UMC_V8_10_PER_CHANNEL_OFFSET;
576 		adev->umc.retire_unit = UMC_V8_10_NA_COL_2BITS_POWER_OF_2_NUM;
577 		if (adev->umc.node_inst_num == 4)
578 			adev->umc.channel_idx_tbl = &umc_v8_10_channel_idx_tbl_ext0[0][0][0];
579 		else
580 			adev->umc.channel_idx_tbl = &umc_v8_10_channel_idx_tbl[0][0][0];
581 		adev->umc.ras = &umc_v8_10_ras;
582 		break;
583 	case IP_VERSION(8, 11, 0):
584 		break;
585 	default:
586 		break;
587 	}
588 }
589 
590 
591 static void gmc_v11_0_set_mmhub_funcs(struct amdgpu_device *adev)
592 {
593 	switch (adev->ip_versions[MMHUB_HWIP][0]) {
594 	case IP_VERSION(3, 0, 1):
595 		adev->mmhub.funcs = &mmhub_v3_0_1_funcs;
596 		break;
597 	case IP_VERSION(3, 0, 2):
598 		adev->mmhub.funcs = &mmhub_v3_0_2_funcs;
599 		break;
600 	default:
601 		adev->mmhub.funcs = &mmhub_v3_0_funcs;
602 		break;
603 	}
604 }
605 
606 static void gmc_v11_0_set_gfxhub_funcs(struct amdgpu_device *adev)
607 {
608 	switch (adev->ip_versions[GC_HWIP][0]) {
609 	case IP_VERSION(11, 0, 3):
610 		adev->gfxhub.funcs = &gfxhub_v3_0_3_funcs;
611 		break;
612 	default:
613 		adev->gfxhub.funcs = &gfxhub_v3_0_funcs;
614 		break;
615 	}
616 }
617 
618 static int gmc_v11_0_early_init(void *handle)
619 {
620 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
621 
622 	gmc_v11_0_set_gfxhub_funcs(adev);
623 	gmc_v11_0_set_mmhub_funcs(adev);
624 	gmc_v11_0_set_gmc_funcs(adev);
625 	gmc_v11_0_set_irq_funcs(adev);
626 	gmc_v11_0_set_umc_funcs(adev);
627 
628 	adev->gmc.shared_aperture_start = 0x2000000000000000ULL;
629 	adev->gmc.shared_aperture_end =
630 		adev->gmc.shared_aperture_start + (4ULL << 30) - 1;
631 	adev->gmc.private_aperture_start = 0x1000000000000000ULL;
632 	adev->gmc.private_aperture_end =
633 		adev->gmc.private_aperture_start + (4ULL << 30) - 1;
634 
635 	return 0;
636 }
637 
638 static int gmc_v11_0_late_init(void *handle)
639 {
640 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
641 	int r;
642 
643 	r = amdgpu_gmc_allocate_vm_inv_eng(adev);
644 	if (r)
645 		return r;
646 
647 	r = amdgpu_gmc_ras_late_init(adev);
648 	if (r)
649 		return r;
650 
651 	return amdgpu_irq_get(adev, &adev->gmc.vm_fault, 0);
652 }
653 
654 static void gmc_v11_0_vram_gtt_location(struct amdgpu_device *adev,
655 					struct amdgpu_gmc *mc)
656 {
657 	u64 base = 0;
658 
659 	base = adev->mmhub.funcs->get_fb_location(adev);
660 
661 	amdgpu_gmc_vram_location(adev, &adev->gmc, base);
662 	amdgpu_gmc_gart_location(adev, mc);
663 	amdgpu_gmc_agp_location(adev, mc);
664 
665 	/* base offset of vram pages */
666 	if (amdgpu_sriov_vf(adev))
667 		adev->vm_manager.vram_base_offset = 0;
668 	else
669 		adev->vm_manager.vram_base_offset = adev->mmhub.funcs->get_mc_fb_offset(adev);
670 }
671 
672 /**
673  * gmc_v11_0_mc_init - initialize the memory controller driver params
674  *
675  * @adev: amdgpu_device pointer
676  *
677  * Look up the amount of vram, vram width, and decide how to place
678  * vram and gart within the GPU's physical address space.
679  * Returns 0 for success.
680  */
681 static int gmc_v11_0_mc_init(struct amdgpu_device *adev)
682 {
683 	int r;
684 
685 	/* size in MB on si */
686 	adev->gmc.mc_vram_size =
687 		adev->nbio.funcs->get_memsize(adev) * 1024ULL * 1024ULL;
688 	adev->gmc.real_vram_size = adev->gmc.mc_vram_size;
689 
690 	if (!(adev->flags & AMD_IS_APU)) {
691 		r = amdgpu_device_resize_fb_bar(adev);
692 		if (r)
693 			return r;
694 	}
695 	adev->gmc.aper_base = pci_resource_start(adev->pdev, 0);
696 	adev->gmc.aper_size = pci_resource_len(adev->pdev, 0);
697 
698 #ifdef CONFIG_X86_64
699 	if ((adev->flags & AMD_IS_APU) && !amdgpu_passthrough(adev)) {
700 		adev->gmc.aper_base = adev->mmhub.funcs->get_mc_fb_offset(adev);
701 		adev->gmc.aper_size = adev->gmc.real_vram_size;
702 	}
703 #endif
704 	/* In case the PCI BAR is larger than the actual amount of vram */
705 	adev->gmc.visible_vram_size = adev->gmc.aper_size;
706 	if (adev->gmc.visible_vram_size > adev->gmc.real_vram_size)
707 		adev->gmc.visible_vram_size = adev->gmc.real_vram_size;
708 
709 	/* set the gart size */
710 	if (amdgpu_gart_size == -1) {
711 		adev->gmc.gart_size = 512ULL << 20;
712 	} else
713 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
714 
715 	gmc_v11_0_vram_gtt_location(adev, &adev->gmc);
716 
717 	return 0;
718 }
719 
720 static int gmc_v11_0_gart_init(struct amdgpu_device *adev)
721 {
722 	int r;
723 
724 	if (adev->gart.bo) {
725 		WARN(1, "PCIE GART already initialized\n");
726 		return 0;
727 	}
728 
729 	/* Initialize common gart structure */
730 	r = amdgpu_gart_init(adev);
731 	if (r)
732 		return r;
733 
734 	adev->gart.table_size = adev->gart.num_gpu_pages * 8;
735 	adev->gart.gart_pte_flags = AMDGPU_PTE_MTYPE_NV10(MTYPE_UC) |
736 				 AMDGPU_PTE_EXECUTABLE;
737 
738 	return amdgpu_gart_table_vram_alloc(adev);
739 }
740 
741 static int gmc_v11_0_sw_init(void *handle)
742 {
743 	int r, vram_width = 0, vram_type = 0, vram_vendor = 0;
744 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
745 
746 	adev->mmhub.funcs->init(adev);
747 
748 	spin_lock_init(&adev->gmc.invalidate_lock);
749 
750 	r = amdgpu_atomfirmware_get_vram_info(adev,
751 					      &vram_width, &vram_type, &vram_vendor);
752 	adev->gmc.vram_width = vram_width;
753 
754 	adev->gmc.vram_type = vram_type;
755 	adev->gmc.vram_vendor = vram_vendor;
756 
757 	switch (adev->ip_versions[GC_HWIP][0]) {
758 	case IP_VERSION(11, 0, 0):
759 	case IP_VERSION(11, 0, 1):
760 	case IP_VERSION(11, 0, 2):
761 	case IP_VERSION(11, 0, 3):
762 	case IP_VERSION(11, 0, 4):
763 		adev->num_vmhubs = 2;
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 		printk(KERN_WARNING "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(void *handle)
856 {
857 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
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_MMHUB_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_MMHUB_0, 0);
906 
907 	DRM_INFO("PCIE GART of %uM enabled (table at 0x%016llX).\n",
908 		 (unsigned)(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(void *handle)
915 {
916 	int r;
917 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
918 
919 	/* The sequence of these two function calls matters.*/
920 	gmc_v11_0_init_golden_registers(adev);
921 
922 	r = gmc_v11_0_gart_enable(adev);
923 	if (r)
924 		return r;
925 
926 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
927 		adev->umc.funcs->init_registers(adev);
928 
929 	return 0;
930 }
931 
932 /**
933  * gmc_v11_0_gart_disable - gart disable
934  *
935  * @adev: amdgpu_device pointer
936  *
937  * This disables all VM page table.
938  */
939 static void gmc_v11_0_gart_disable(struct amdgpu_device *adev)
940 {
941 	adev->mmhub.funcs->gart_disable(adev);
942 }
943 
944 static int gmc_v11_0_hw_fini(void *handle)
945 {
946 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
947 
948 	if (amdgpu_sriov_vf(adev)) {
949 		/* full access mode, so don't touch any GMC register */
950 		DRM_DEBUG("For SRIOV client, shouldn't do anything.\n");
951 		return 0;
952 	}
953 
954 	amdgpu_irq_put(adev, &adev->gmc.vm_fault, 0);
955 	gmc_v11_0_gart_disable(adev);
956 
957 	return 0;
958 }
959 
960 static int gmc_v11_0_suspend(void *handle)
961 {
962 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
963 
964 	gmc_v11_0_hw_fini(adev);
965 
966 	return 0;
967 }
968 
969 static int gmc_v11_0_resume(void *handle)
970 {
971 	int r;
972 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
973 
974 	r = gmc_v11_0_hw_init(adev);
975 	if (r)
976 		return r;
977 
978 	amdgpu_vmid_reset_all(adev);
979 
980 	return 0;
981 }
982 
983 static bool gmc_v11_0_is_idle(void *handle)
984 {
985 	/* MC is always ready in GMC v11.*/
986 	return true;
987 }
988 
989 static int gmc_v11_0_wait_for_idle(void *handle)
990 {
991 	/* There is no need to wait for MC idle in GMC v11.*/
992 	return 0;
993 }
994 
995 static int gmc_v11_0_soft_reset(void *handle)
996 {
997 	return 0;
998 }
999 
1000 static int gmc_v11_0_set_clockgating_state(void *handle,
1001 					   enum amd_clockgating_state state)
1002 {
1003 	int r;
1004 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1005 
1006 	r = adev->mmhub.funcs->set_clockgating(adev, state);
1007 	if (r)
1008 		return r;
1009 
1010 	return athub_v3_0_set_clockgating(adev, state);
1011 }
1012 
1013 static void gmc_v11_0_get_clockgating_state(void *handle, u64 *flags)
1014 {
1015 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
1016 
1017 	adev->mmhub.funcs->get_clockgating(adev, flags);
1018 
1019 	athub_v3_0_get_clockgating(adev, flags);
1020 }
1021 
1022 static int gmc_v11_0_set_powergating_state(void *handle,
1023 					   enum amd_powergating_state state)
1024 {
1025 	return 0;
1026 }
1027 
1028 const struct amd_ip_funcs gmc_v11_0_ip_funcs = {
1029 	.name = "gmc_v11_0",
1030 	.early_init = gmc_v11_0_early_init,
1031 	.sw_init = gmc_v11_0_sw_init,
1032 	.hw_init = gmc_v11_0_hw_init,
1033 	.late_init = gmc_v11_0_late_init,
1034 	.sw_fini = gmc_v11_0_sw_fini,
1035 	.hw_fini = gmc_v11_0_hw_fini,
1036 	.suspend = gmc_v11_0_suspend,
1037 	.resume = gmc_v11_0_resume,
1038 	.is_idle = gmc_v11_0_is_idle,
1039 	.wait_for_idle = gmc_v11_0_wait_for_idle,
1040 	.soft_reset = gmc_v11_0_soft_reset,
1041 	.set_clockgating_state = gmc_v11_0_set_clockgating_state,
1042 	.set_powergating_state = gmc_v11_0_set_powergating_state,
1043 	.get_clockgating_state = gmc_v11_0_get_clockgating_state,
1044 };
1045 
1046 const struct amdgpu_ip_block_version gmc_v11_0_ip_block = {
1047 	.type = AMD_IP_BLOCK_TYPE_GMC,
1048 	.major = 11,
1049 	.minor = 0,
1050 	.rev = 0,
1051 	.funcs = &gmc_v11_0_ip_funcs,
1052 };
1053