xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c (revision c06b6cde2a1c3bcbb561bd57bb6f34eae9030921)
1 /*
2  * Copyright 2018 Advanced Micro Devices, Inc.
3  * All Rights Reserved.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the
7  * "Software"), to deal in the Software without restriction, including
8  * without limitation the rights to use, copy, modify, merge, publish,
9  * distribute, sub license, and/or sell copies of the Software, and to
10  * permit persons to whom the Software is furnished to do so, subject to
11  * the following conditions:
12  *
13  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
14  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
15  * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
16  * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
17  * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
18  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
19  * USE OR OTHER DEALINGS IN THE SOFTWARE.
20  *
21  * The above copyright notice and this permission notice (including the
22  * next paragraph) shall be included in all copies or substantial portions
23  * of the Software.
24  *
25  */
26 
27 #include <linux/io-64-nonatomic-lo-hi.h>
28 #ifdef CONFIG_X86
29 #include <asm/hypervisor.h>
30 #endif
31 
32 #include "amdgpu.h"
33 #include "amdgpu_gmc.h"
34 #include "amdgpu_ras.h"
35 #include "amdgpu_reset.h"
36 #include "amdgpu_xgmi.h"
37 #include "amdgpu_atomfirmware.h"
38 
39 #include <drm/drm_drv.h>
40 #include <drm/ttm/ttm_tt.h>
41 
42 static const u64 four_gb = 0x100000000ULL;
43 
44 bool amdgpu_gmc_is_pdb0_enabled(struct amdgpu_device *adev)
45 {
46 	return adev->gmc.xgmi.connected_to_cpu || amdgpu_virt_xgmi_migrate_enabled(adev);
47 }
48 
49 /**
50  * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
51  *
52  * @adev: amdgpu_device pointer
53  *
54  * Allocate video memory for pdb0 and map it for CPU access
55  * Returns 0 for success, error for failure.
56  */
57 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
58 {
59 	int r;
60 	struct amdgpu_bo_param bp;
61 	u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
62 	uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
63 	uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) - 1) >> pde0_page_shift;
64 
65 	memset(&bp, 0, sizeof(bp));
66 	bp.size = PAGE_ALIGN((npdes + 1) * 8);
67 	bp.byte_align = PAGE_SIZE;
68 	bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
69 	bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
70 		AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
71 	bp.type = ttm_bo_type_kernel;
72 	bp.resv = NULL;
73 	bp.bo_ptr_size = sizeof(struct amdgpu_bo);
74 
75 	r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
76 	if (r)
77 		return r;
78 
79 	r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
80 	if (unlikely(r != 0))
81 		goto bo_reserve_failure;
82 
83 	r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
84 	if (r)
85 		goto bo_pin_failure;
86 	r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
87 	if (r)
88 		goto bo_kmap_failure;
89 
90 	amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
91 	return 0;
92 
93 bo_kmap_failure:
94 	amdgpu_bo_unpin(adev->gmc.pdb0_bo);
95 bo_pin_failure:
96 	amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
97 bo_reserve_failure:
98 	amdgpu_bo_unref(&adev->gmc.pdb0_bo);
99 	return r;
100 }
101 
102 /**
103  * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
104  *
105  * @bo: the BO to get the PDE for
106  * @level: the level in the PD hirarchy
107  * @addr: resulting addr
108  * @flags: resulting flags
109  *
110  * Get the address and flags to be used for a PDE (Page Directory Entry).
111  */
112 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
113 			       uint64_t *addr, uint64_t *flags)
114 {
115 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
116 
117 	switch (bo->tbo.resource->mem_type) {
118 	case TTM_PL_TT:
119 		*addr = bo->tbo.ttm->dma_address[0];
120 		break;
121 	case TTM_PL_VRAM:
122 		*addr = amdgpu_bo_gpu_offset(bo);
123 		break;
124 	default:
125 		*addr = 0;
126 		break;
127 	}
128 	*flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource);
129 	amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
130 }
131 
132 /*
133  * amdgpu_gmc_pd_addr - return the address of the root directory
134  */
135 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
136 {
137 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
138 	uint64_t pd_addr;
139 
140 	/* TODO: move that into ASIC specific code */
141 	if (adev->asic_type >= CHIP_VEGA10) {
142 		uint64_t flags = AMDGPU_PTE_VALID;
143 
144 		amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
145 		pd_addr |= flags;
146 	} else {
147 		pd_addr = amdgpu_bo_gpu_offset(bo);
148 	}
149 	return pd_addr;
150 }
151 
152 /**
153  * amdgpu_gmc_set_pte_pde - update the page tables using CPU
154  *
155  * @adev: amdgpu_device pointer
156  * @cpu_pt_addr: cpu address of the page table
157  * @gpu_page_idx: entry in the page table to update
158  * @addr: dst addr to write into pte/pde
159  * @flags: access flags
160  *
161  * Update the page tables using CPU.
162  */
163 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
164 				uint32_t gpu_page_idx, uint64_t addr,
165 				uint64_t flags)
166 {
167 	void __iomem *ptr = (void *)cpu_pt_addr;
168 	uint64_t value;
169 
170 	/*
171 	 * The following is for PTE only. GART does not have PDEs.
172 	*/
173 	value = addr & 0x0000FFFFFFFFF000ULL;
174 	value |= flags;
175 	writeq(value, ptr + (gpu_page_idx * 8));
176 
177 	return 0;
178 }
179 
180 /**
181  * amdgpu_gmc_agp_addr - return the address in the AGP address space
182  *
183  * @bo: TTM BO which needs the address, must be in GTT domain
184  *
185  * Tries to figure out how to access the BO through the AGP aperture. Returns
186  * AMDGPU_BO_INVALID_OFFSET if that is not possible.
187  */
188 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
189 {
190 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
191 
192 	if (!bo->ttm)
193 		return AMDGPU_BO_INVALID_OFFSET;
194 
195 	if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
196 		return AMDGPU_BO_INVALID_OFFSET;
197 
198 	if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
199 		return AMDGPU_BO_INVALID_OFFSET;
200 
201 	return adev->gmc.agp_start + bo->ttm->dma_address[0];
202 }
203 
204 /**
205  * amdgpu_gmc_vram_location - try to find VRAM location
206  *
207  * @adev: amdgpu device structure holding all necessary information
208  * @mc: memory controller structure holding memory information
209  * @base: base address at which to put VRAM
210  *
211  * Function will try to place VRAM at base address provided
212  * as parameter.
213  */
214 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
215 			      u64 base)
216 {
217 	uint64_t vis_limit = (uint64_t)amdgpu_vis_vram_limit << 20;
218 	uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
219 
220 	mc->vram_start = base;
221 	mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
222 	if (limit < mc->real_vram_size)
223 		mc->real_vram_size = limit;
224 
225 	if (vis_limit && vis_limit < mc->visible_vram_size)
226 		mc->visible_vram_size = vis_limit;
227 
228 	if (mc->real_vram_size < mc->visible_vram_size)
229 		mc->visible_vram_size = mc->real_vram_size;
230 
231 	if (mc->xgmi.num_physical_nodes == 0) {
232 		mc->fb_start = mc->vram_start;
233 		mc->fb_end = mc->vram_end;
234 	}
235 	dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
236 			mc->mc_vram_size >> 20, mc->vram_start,
237 			mc->vram_end, mc->real_vram_size >> 20);
238 }
239 
240 /** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
241  *
242  * @adev: amdgpu device structure holding all necessary information
243  * @mc: memory controller structure holding memory information
244  *
245  * This function is only used if use GART for FB translation. In such
246  * case, we use sysvm aperture (vmid0 page tables) for both vram
247  * and gart (aka system memory) access.
248  *
249  * GPUVM (and our organization of vmid0 page tables) require sysvm
250  * aperture to be placed at a location aligned with 8 times of native
251  * page size. For example, if vm_context0_cntl.page_table_block_size
252  * is 12, then native page size is 8G (2M*2^12), sysvm should start
253  * with a 64G aligned address. For simplicity, we just put sysvm at
254  * address 0. So vram start at address 0 and gart is right after vram.
255  */
256 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
257 {
258 	u64 hive_vram_start = 0;
259 	u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
260 	mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
261 	mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
262 	/* node_segment_size may not 4GB aligned on SRIOV, align up is needed. */
263 	mc->gart_start = ALIGN(hive_vram_end + 1, four_gb);
264 	mc->gart_end = mc->gart_start + mc->gart_size - 1;
265 	if (amdgpu_virt_xgmi_migrate_enabled(adev)) {
266 		/* set mc->vram_start to 0 to switch the returned GPU address of
267 		 * amdgpu_bo_create_reserved() from FB aperture to GART aperture.
268 		 */
269 		mc->vram_start = 0;
270 		mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
271 		mc->visible_vram_size = min(mc->visible_vram_size, mc->real_vram_size);
272 	} else {
273 		mc->fb_start = hive_vram_start;
274 		mc->fb_end = hive_vram_end;
275 	}
276 	dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
277 			mc->mc_vram_size >> 20, mc->vram_start,
278 			mc->vram_end, mc->real_vram_size >> 20);
279 	dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
280 			mc->gart_size >> 20, mc->gart_start, mc->gart_end);
281 }
282 
283 void amdgpu_gmc_set_gart_size(struct amdgpu_device *adev, u64 default_size)
284 {
285 	if (amdgpu_gart_size == -1)
286 		adev->gmc.gart_size =
287 			default_size + adev->pm.smu_prv_buffer_size;
288 	else
289 		adev->gmc.gart_size = (u64)amdgpu_gart_size << 20;
290 }
291 
292 /**
293  * amdgpu_gmc_gart_location - try to find GART location
294  *
295  * @adev: amdgpu device structure holding all necessary information
296  * @mc: memory controller structure holding memory information
297  * @gart_placement: GART placement policy with respect to VRAM
298  *
299  * Function will try to place GART before or after VRAM.
300  * If GART size is bigger than space left then we ajust GART size.
301  * Thus function will never fails.
302  */
303 void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
304 			      enum amdgpu_gart_placement gart_placement)
305 {
306 	u64 size_af, size_bf;
307 	/*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
308 	u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
309 
310 	/* VCE doesn't like it when BOs cross a 4GB segment, so align
311 	 * the GART base on a 4GB boundary as well.
312 	 */
313 	size_bf = mc->fb_start;
314 	size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
315 
316 	if (mc->gart_size > max(size_bf, size_af)) {
317 		dev_warn(adev->dev, "limiting GART\n");
318 		mc->gart_size = max(size_bf, size_af);
319 	}
320 
321 	switch (gart_placement) {
322 	case AMDGPU_GART_PLACEMENT_HIGH:
323 		mc->gart_start = max_mc_address - mc->gart_size + 1;
324 		break;
325 	case AMDGPU_GART_PLACEMENT_LOW:
326 		if (size_bf >= mc->gart_size)
327 			mc->gart_start = 0;
328 		else
329 			mc->gart_start = ALIGN(mc->fb_end, four_gb);
330 		break;
331 	case AMDGPU_GART_PLACEMENT_BEST_FIT:
332 	default:
333 		if ((size_bf >= mc->gart_size && size_bf < size_af) ||
334 		    (size_af < mc->gart_size))
335 			mc->gart_start = 0;
336 		else
337 			mc->gart_start = max_mc_address - mc->gart_size + 1;
338 		break;
339 	}
340 
341 	mc->gart_start &= ~(four_gb - 1);
342 	mc->gart_end = mc->gart_start + mc->gart_size - 1;
343 	dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
344 			mc->gart_size >> 20, mc->gart_start, mc->gart_end);
345 }
346 
347 /**
348  * amdgpu_gmc_agp_location - try to find AGP location
349  * @adev: amdgpu device structure holding all necessary information
350  * @mc: memory controller structure holding memory information
351  *
352  * Function will place try to find a place for the AGP BAR in the MC address
353  * space.
354  *
355  * AGP BAR will be assigned the largest available hole in the address space.
356  * Should be called after VRAM and GART locations are setup.
357  */
358 void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
359 {
360 	const uint64_t sixteen_gb = 1ULL << 34;
361 	const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
362 	u64 size_af, size_bf;
363 
364 	if (mc->fb_start > mc->gart_start) {
365 		size_bf = (mc->fb_start & sixteen_gb_mask) -
366 			ALIGN(mc->gart_end + 1, sixteen_gb);
367 		size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
368 	} else {
369 		size_bf = mc->fb_start & sixteen_gb_mask;
370 		size_af = (mc->gart_start & sixteen_gb_mask) -
371 			ALIGN(mc->fb_end + 1, sixteen_gb);
372 	}
373 
374 	if (size_bf > size_af) {
375 		mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
376 		mc->agp_size = size_bf;
377 	} else {
378 		mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
379 		mc->agp_size = size_af;
380 	}
381 
382 	mc->agp_end = mc->agp_start + mc->agp_size - 1;
383 	dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
384 			mc->agp_size >> 20, mc->agp_start, mc->agp_end);
385 }
386 
387 /**
388  * amdgpu_gmc_set_agp_default - Set the default AGP aperture value.
389  * @adev: amdgpu device structure holding all necessary information
390  * @mc: memory controller structure holding memory information
391  *
392  * To disable the AGP aperture, you need to set the start to a larger
393  * value than the end.  This function sets the default value which
394  * can then be overridden using amdgpu_gmc_agp_location() if you want
395  * to enable the AGP aperture on a specific chip.
396  *
397  */
398 void amdgpu_gmc_set_agp_default(struct amdgpu_device *adev,
399 				struct amdgpu_gmc *mc)
400 {
401 	mc->agp_start = 0xffffffffffff;
402 	mc->agp_end = 0;
403 	mc->agp_size = 0;
404 }
405 
406 /**
407  * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid
408  *
409  * @addr: 48 bit physical address, page aligned (36 significant bits)
410  * @pasid: 16 bit process address space identifier
411  */
412 static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid)
413 {
414 	return addr << 4 | pasid;
415 }
416 
417 /**
418  * amdgpu_gmc_filter_faults - filter VM faults
419  *
420  * @adev: amdgpu device structure
421  * @ih: interrupt ring that the fault received from
422  * @addr: address of the VM fault
423  * @pasid: PASID of the process causing the fault
424  * @timestamp: timestamp of the fault
425  *
426  * Returns:
427  * True if the fault was filtered and should not be processed further.
428  * False if the fault is a new one and needs to be handled.
429  */
430 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
431 			      struct amdgpu_ih_ring *ih, uint64_t addr,
432 			      uint16_t pasid, uint64_t timestamp)
433 {
434 	struct amdgpu_gmc *gmc = &adev->gmc;
435 	uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid);
436 	struct amdgpu_gmc_fault *fault;
437 	uint32_t hash;
438 
439 	/* Stale retry fault if timestamp goes backward */
440 	if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp))
441 		return true;
442 
443 	/* If we don't have space left in the ring buffer return immediately */
444 	stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
445 		AMDGPU_GMC_FAULT_TIMEOUT;
446 	if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
447 		return true;
448 
449 	/* Try to find the fault in the hash */
450 	hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
451 	fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
452 	while (fault->timestamp >= stamp) {
453 		uint64_t tmp;
454 
455 		if (atomic64_read(&fault->key) == key) {
456 			/*
457 			 * if we get a fault which is already present in
458 			 * the fault_ring and the timestamp of
459 			 * the fault is after the expired timestamp,
460 			 * then this is a new fault that needs to be added
461 			 * into the fault ring.
462 			 */
463 			if (fault->timestamp_expiry != 0 &&
464 			    amdgpu_ih_ts_after(fault->timestamp_expiry,
465 					       timestamp))
466 				break;
467 			else
468 				return true;
469 		}
470 
471 		tmp = fault->timestamp;
472 		fault = &gmc->fault_ring[fault->next];
473 
474 		/* Check if the entry was reused */
475 		if (fault->timestamp >= tmp)
476 			break;
477 	}
478 
479 	/* Add the fault to the ring */
480 	fault = &gmc->fault_ring[gmc->last_fault];
481 	atomic64_set(&fault->key, key);
482 	fault->timestamp = timestamp;
483 
484 	/* And update the hash */
485 	fault->next = gmc->fault_hash[hash].idx;
486 	gmc->fault_hash[hash].idx = gmc->last_fault++;
487 	return false;
488 }
489 
490 /**
491  * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter
492  *
493  * @adev: amdgpu device structure
494  * @addr: address of the VM fault
495  * @pasid: PASID of the process causing the fault
496  *
497  * Remove the address from fault filter, then future vm fault on this address
498  * will pass to retry fault handler to recover.
499  */
500 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
501 				     uint16_t pasid)
502 {
503 	struct amdgpu_gmc *gmc = &adev->gmc;
504 	uint64_t key = amdgpu_gmc_fault_key(addr, pasid);
505 	struct amdgpu_ih_ring *ih;
506 	struct amdgpu_gmc_fault *fault;
507 	uint32_t last_wptr;
508 	uint64_t last_ts;
509 	uint32_t hash;
510 	uint64_t tmp;
511 
512 	if (adev->irq.retry_cam_enabled)
513 		return;
514 	else if (adev->irq.ih1.ring_size)
515 		ih = &adev->irq.ih1;
516 	else if (adev->irq.ih_soft.enabled)
517 		ih = &adev->irq.ih_soft;
518 	else
519 		return;
520 
521 	/* Get the WPTR of the last entry in IH ring */
522 	last_wptr = amdgpu_ih_get_wptr(adev, ih);
523 	/* Order wptr with ring data. */
524 	rmb();
525 	/* Get the timetamp of the last entry in IH ring */
526 	last_ts = amdgpu_ih_decode_iv_ts(adev, ih, last_wptr, -1);
527 
528 	hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
529 	fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
530 	do {
531 		if (atomic64_read(&fault->key) == key) {
532 			/*
533 			 * Update the timestamp when this fault
534 			 * expired.
535 			 */
536 			fault->timestamp_expiry = last_ts;
537 			break;
538 		}
539 
540 		tmp = fault->timestamp;
541 		fault = &gmc->fault_ring[fault->next];
542 	} while (fault->timestamp < tmp);
543 }
544 
545 int amdgpu_gmc_handle_retry_fault(struct amdgpu_device *adev,
546 				  struct amdgpu_iv_entry *entry,
547 				  u64 addr,
548 				  u32 cam_index,
549 				  u32 node_id,
550 				  bool write_fault)
551 {
552 	int ret;
553 
554 	if (adev->irq.retry_cam_enabled) {
555 		/* Delegate it to a different ring if the hardware hasn't
556 		 * already done it.
557 		 */
558 		if (entry->ih == &adev->irq.ih) {
559 			amdgpu_irq_delegate(adev, entry, 8);
560 			return 1;
561 		}
562 
563 		ret = amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id,
564 					     addr, entry->timestamp, write_fault);
565 		WDOORBELL32(adev->irq.retry_cam_doorbell_index, cam_index);
566 		if (ret)
567 			return 1;
568 	} else {
569 		/* Process it only if it's the first fault for this address */
570 		if (entry->ih != &adev->irq.ih_soft &&
571 		    amdgpu_gmc_filter_faults(adev, entry->ih, addr, entry->pasid,
572 					     entry->timestamp))
573 			return 1;
574 
575 		/* Delegate it to a different ring if the hardware hasn't
576 		 * already done it.
577 		 */
578 		if (entry->ih == &adev->irq.ih) {
579 			amdgpu_irq_delegate(adev, entry, 8);
580 			return 1;
581 		}
582 
583 		/* Try to handle the recoverable page faults by filling page
584 		 * tables
585 		 */
586 		if (amdgpu_vm_handle_fault(adev, entry->pasid, entry->vmid, node_id,
587 					   addr, entry->timestamp, write_fault))
588 			return 1;
589 	}
590 	return 0;
591 }
592 
593 int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev)
594 {
595 	int r;
596 
597 	/* umc ras block */
598 	r = amdgpu_umc_ras_sw_init(adev);
599 	if (r)
600 		return r;
601 
602 	/* mmhub ras block */
603 	r = amdgpu_mmhub_ras_sw_init(adev);
604 	if (r)
605 		return r;
606 
607 	/* hdp ras block */
608 	r = amdgpu_hdp_ras_sw_init(adev);
609 	if (r)
610 		return r;
611 
612 	/* mca.x ras block */
613 	r = amdgpu_mca_mp0_ras_sw_init(adev);
614 	if (r)
615 		return r;
616 
617 	r = amdgpu_mca_mp1_ras_sw_init(adev);
618 	if (r)
619 		return r;
620 
621 	r = amdgpu_mca_mpio_ras_sw_init(adev);
622 	if (r)
623 		return r;
624 
625 	/* xgmi ras block */
626 	r = amdgpu_xgmi_ras_sw_init(adev);
627 	if (r)
628 		return r;
629 
630 	return 0;
631 }
632 
633 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
634 {
635 	return 0;
636 }
637 
638 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
639 {
640 
641 }
642 
643 	/*
644 	 * The latest engine allocation on gfx9/10 is:
645 	 * Engine 2, 3: firmware
646 	 * Engine 0, 1, 4~16: amdgpu ring,
647 	 *                    subject to change when ring number changes
648 	 * Engine 17: Gart flushes
649 	 */
650 #define AMDGPU_VMHUB_INV_ENG_BITMAP		0x1FFF3
651 
652 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
653 {
654 	struct amdgpu_ring *ring;
655 	unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] = {0};
656 	unsigned i;
657 	unsigned vmhub, inv_eng;
658 	struct amdgpu_ring *shared_ring;
659 
660 	/* init the vm inv eng for all vmhubs */
661 	for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
662 		vm_inv_engs[i] = AMDGPU_VMHUB_INV_ENG_BITMAP;
663 		/* reserve engine 5 for firmware */
664 		if (adev->enable_mes)
665 			vm_inv_engs[i] &= ~(1 << 5);
666 		/* reserve engine 6 for uni mes */
667 		if (adev->enable_uni_mes)
668 			vm_inv_engs[i] &= ~(1 << 6);
669 		/* reserve mmhub engine 3 for firmware */
670 		if (adev->enable_umsch_mm)
671 			vm_inv_engs[i] &= ~(1 << 3);
672 	}
673 
674 	for (i = 0; i < adev->num_rings; ++i) {
675 		ring = adev->rings[i];
676 		vmhub = ring->vm_hub;
677 
678 		if (ring == &adev->mes.ring[0] ||
679 		    ring == &adev->mes.ring[1] ||
680 		    ring == &adev->umsch_mm.ring ||
681 		    ring == &adev->cper.ring_buf)
682 			continue;
683 
684 		/* Skip if the ring is a shared ring */
685 		if (amdgpu_sdma_is_shared_inv_eng(adev, ring))
686 			continue;
687 
688 		inv_eng = ffs(vm_inv_engs[vmhub]);
689 		if (!inv_eng) {
690 			dev_err(adev->dev, "no VM inv eng for ring %s\n",
691 				ring->name);
692 			return -EINVAL;
693 		}
694 
695 		ring->vm_inv_eng = inv_eng - 1;
696 		vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
697 
698 		dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
699 			 ring->name, ring->vm_inv_eng, ring->vm_hub);
700 		/* SDMA has a special packet which allows it to use the same
701 		 * invalidation engine for all the rings in one instance.
702 		 * Therefore, we do not allocate a separate VM invalidation engine
703 		 * for SDMA page rings. Instead, they share the VM invalidation
704 		 * engine with the SDMA gfx ring. This change ensures efficient
705 		 * resource management and avoids the issue of insufficient VM
706 		 * invalidation engines.
707 		 */
708 		shared_ring = amdgpu_sdma_get_shared_ring(adev, ring);
709 		if (shared_ring) {
710 			shared_ring->vm_inv_eng = ring->vm_inv_eng;
711 			dev_info(adev->dev, "ring %s shares VM invalidation engine %u with ring %s on hub %u\n",
712 					ring->name, ring->vm_inv_eng, shared_ring->name, ring->vm_hub);
713 			continue;
714 		}
715 	}
716 
717 	return 0;
718 }
719 
720 void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
721 			      uint32_t vmhub, uint32_t flush_type)
722 {
723 	struct amdgpu_ring *ring;
724 	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
725 	struct dma_fence *fence;
726 	struct amdgpu_job *job;
727 	int r;
728 
729 	ring = to_amdgpu_ring(adev->mman.buffer_funcs_scheds[0]);
730 
731 	if (!hub->sdma_invalidation_workaround || vmid ||
732 	    !adev->mman.buffer_funcs_enabled || !adev->ib_pool_ready ||
733 	    !ring->sched.ready) {
734 		/*
735 		 * A GPU reset should flush all TLBs anyway, so no need to do
736 		 * this while one is ongoing.
737 		 */
738 		if (!down_read_trylock(&adev->reset_domain->sem))
739 			return;
740 
741 		if (adev->gmc.flush_tlb_needs_extra_type_2)
742 			adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid,
743 							   vmhub, 2);
744 
745 		if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2)
746 			adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid,
747 							   vmhub, 0);
748 
749 		adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid, vmhub,
750 						   flush_type);
751 		up_read(&adev->reset_domain->sem);
752 		return;
753 	}
754 
755 	/* The SDMA on Navi 1x has a bug which can theoretically result in memory
756 	 * corruption if an invalidation happens at the same time as an VA
757 	 * translation. Avoid this by doing the invalidation from the SDMA
758 	 * itself at least for GART.
759 	 */
760 	mutex_lock(&adev->mman.default_entity.lock);
761 	r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.default_entity.base,
762 				     AMDGPU_FENCE_OWNER_UNDEFINED,
763 				     16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
764 				     &job, AMDGPU_KERNEL_JOB_ID_FLUSH_GPU_TLB);
765 	if (r)
766 		goto error_alloc;
767 
768 	job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
769 	job->vm_needs_flush = true;
770 	job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
771 	amdgpu_ring_pad_ib(ring, &job->ibs[0]);
772 	fence = amdgpu_job_submit(job);
773 	mutex_unlock(&adev->mman.default_entity.lock);
774 
775 	dma_fence_wait(fence, false);
776 	dma_fence_put(fence);
777 
778 	return;
779 
780 error_alloc:
781 	mutex_unlock(&adev->mman.default_entity.lock);
782 	dev_err(adev->dev, "Error flushing GPU TLB using the SDMA (%d)!\n", r);
783 }
784 
785 int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid,
786 				   uint32_t flush_type, bool all_hub,
787 				   uint32_t inst)
788 {
789 	struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring;
790 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst];
791 	unsigned int ndw;
792 	int r, cnt = 0;
793 	uint32_t seq;
794 
795 	/*
796 	 * A GPU reset should flush all TLBs anyway, so no need to do
797 	 * this while one is ongoing.
798 	 */
799 	if (!down_read_trylock(&adev->reset_domain->sem))
800 		return 0;
801 
802 	if (!adev->gmc.flush_pasid_uses_kiq || !ring->sched.ready) {
803 
804 		if (!adev->gmc.gmc_funcs->flush_gpu_tlb_pasid) {
805 			r = 0;
806 			goto error_unlock_reset;
807 		}
808 
809 		if (adev->gmc.flush_tlb_needs_extra_type_2)
810 			adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
811 								 2, all_hub,
812 								 inst);
813 
814 		if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2)
815 			adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
816 								 0, all_hub,
817 								 inst);
818 
819 		adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
820 							 flush_type, all_hub,
821 							 inst);
822 		r = 0;
823 	} else {
824 		/* 2 dwords flush + 8 dwords fence */
825 		ndw = kiq->pmf->invalidate_tlbs_size + 8;
826 
827 		if (adev->gmc.flush_tlb_needs_extra_type_2)
828 			ndw += kiq->pmf->invalidate_tlbs_size;
829 
830 		if (adev->gmc.flush_tlb_needs_extra_type_0)
831 			ndw += kiq->pmf->invalidate_tlbs_size;
832 
833 		spin_lock(&adev->gfx.kiq[inst].ring_lock);
834 		r = amdgpu_ring_alloc(ring, ndw);
835 		if (r) {
836 			spin_unlock(&adev->gfx.kiq[inst].ring_lock);
837 			goto error_unlock_reset;
838 		}
839 		if (adev->gmc.flush_tlb_needs_extra_type_2)
840 			kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 2, all_hub);
841 
842 		if (flush_type == 2 && adev->gmc.flush_tlb_needs_extra_type_0)
843 			kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 0, all_hub);
844 
845 		kiq->pmf->kiq_invalidate_tlbs(ring, pasid, flush_type, all_hub);
846 		r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
847 		if (r) {
848 			amdgpu_ring_undo(ring);
849 			spin_unlock(&adev->gfx.kiq[inst].ring_lock);
850 			goto error_unlock_reset;
851 		}
852 
853 		amdgpu_ring_commit(ring);
854 		spin_unlock(&adev->gfx.kiq[inst].ring_lock);
855 
856 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
857 
858 		might_sleep();
859 		while (r < 1 && cnt++ < MAX_KIQ_REG_TRY &&
860 		       !amdgpu_reset_pending(adev->reset_domain)) {
861 			msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
862 			r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
863 		}
864 
865 		if (cnt > MAX_KIQ_REG_TRY) {
866 			dev_err(adev->dev, "timeout waiting for kiq fence\n");
867 			r = -ETIME;
868 		} else
869 			r = 0;
870 	}
871 
872 error_unlock_reset:
873 	up_read(&adev->reset_domain->sem);
874 	return r;
875 }
876 
877 void amdgpu_gmc_fw_reg_write_reg_wait(struct amdgpu_device *adev,
878 				      uint32_t reg0, uint32_t reg1,
879 				      uint32_t ref, uint32_t mask,
880 				      uint32_t xcc_inst)
881 {
882 	struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_inst];
883 	struct amdgpu_ring *ring = &kiq->ring;
884 	signed long r, cnt = 0;
885 	unsigned long flags;
886 	uint32_t seq;
887 
888 	if (adev->mes.ring[MES_PIPE_INST(xcc_inst, 0)].sched.ready) {
889 		amdgpu_mes_reg_write_reg_wait(adev, reg0, reg1,
890 					      ref, mask, xcc_inst);
891 		return;
892 	}
893 
894 	spin_lock_irqsave(&kiq->ring_lock, flags);
895 	amdgpu_ring_alloc(ring, 32);
896 	amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1,
897 					    ref, mask);
898 	r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
899 	if (r)
900 		goto failed_undo;
901 
902 	amdgpu_ring_commit(ring);
903 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
904 
905 	r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
906 
907 	/* don't wait anymore for IRQ context */
908 	if (r < 1 && in_interrupt())
909 		goto failed_kiq;
910 
911 	might_sleep();
912 	while (r < 1 && cnt++ < MAX_KIQ_REG_TRY &&
913 	       !amdgpu_reset_pending(adev->reset_domain)) {
914 
915 		msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
916 		r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
917 	}
918 
919 	if (cnt > MAX_KIQ_REG_TRY)
920 		goto failed_kiq;
921 
922 	return;
923 
924 failed_undo:
925 	amdgpu_ring_undo(ring);
926 	spin_unlock_irqrestore(&kiq->ring_lock, flags);
927 failed_kiq:
928 	dev_err(adev->dev, "failed to write reg %x wait reg %x\n", reg0, reg1);
929 }
930 
931 /**
932  * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
933  * @adev: amdgpu_device pointer
934  *
935  * Check and set if an the device @adev supports Trusted Memory
936  * Zones (TMZ).
937  */
938 void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
939 {
940 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
941 	/* RAVEN */
942 	case IP_VERSION(9, 2, 2):
943 	case IP_VERSION(9, 1, 0):
944 	/* RENOIR looks like RAVEN */
945 	case IP_VERSION(9, 3, 0):
946 	/* GC 10.3.7 */
947 	case IP_VERSION(10, 3, 7):
948 	/* GC 11.0.1 */
949 	case IP_VERSION(11, 0, 1):
950 		if (amdgpu_tmz == 0) {
951 			adev->gmc.tmz_enabled = false;
952 			dev_info(adev->dev,
953 				 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
954 		} else {
955 			adev->gmc.tmz_enabled = true;
956 			dev_info(adev->dev,
957 				 "Trusted Memory Zone (TMZ) feature enabled\n");
958 		}
959 		break;
960 	case IP_VERSION(10, 1, 10):
961 	case IP_VERSION(10, 1, 1):
962 	case IP_VERSION(10, 1, 2):
963 	case IP_VERSION(10, 1, 3):
964 	case IP_VERSION(10, 3, 0):
965 	case IP_VERSION(10, 3, 2):
966 	case IP_VERSION(10, 3, 4):
967 	case IP_VERSION(10, 3, 5):
968 	case IP_VERSION(10, 3, 6):
969 	/* VANGOGH */
970 	case IP_VERSION(10, 3, 1):
971 	/* YELLOW_CARP*/
972 	case IP_VERSION(10, 3, 3):
973 	case IP_VERSION(11, 0, 4):
974 	case IP_VERSION(11, 5, 0):
975 	case IP_VERSION(11, 5, 1):
976 	case IP_VERSION(11, 5, 2):
977 	case IP_VERSION(11, 5, 3):
978 	case IP_VERSION(11, 5, 4):
979 		/* Don't enable it by default yet.
980 		 */
981 		if (amdgpu_tmz < 1) {
982 			adev->gmc.tmz_enabled = false;
983 			dev_info(adev->dev,
984 				 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
985 		} else {
986 			adev->gmc.tmz_enabled = true;
987 			dev_info(adev->dev,
988 				 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
989 		}
990 		break;
991 	default:
992 		adev->gmc.tmz_enabled = false;
993 		dev_info(adev->dev,
994 			 "Trusted Memory Zone (TMZ) feature not supported\n");
995 		break;
996 	}
997 }
998 
999 /**
1000  * amdgpu_gmc_noretry_set -- set per asic noretry defaults
1001  * @adev: amdgpu_device pointer
1002  *
1003  * Set a per asic default for the no-retry parameter.
1004  *
1005  */
1006 void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
1007 {
1008 	struct amdgpu_gmc *gmc = &adev->gmc;
1009 	uint32_t gc_ver = amdgpu_ip_version(adev, GC_HWIP, 0);
1010 	bool noretry_default = (gc_ver == IP_VERSION(9, 0, 1) ||
1011 				gc_ver == IP_VERSION(9, 4, 0) ||
1012 				gc_ver == IP_VERSION(9, 4, 1) ||
1013 				gc_ver == IP_VERSION(9, 4, 2) ||
1014 				gc_ver == IP_VERSION(9, 4, 3) ||
1015 				gc_ver == IP_VERSION(9, 4, 4) ||
1016 				gc_ver == IP_VERSION(9, 5, 0) ||
1017 				gc_ver >= IP_VERSION(10, 3, 0));
1018 
1019 	/* For GFX12.1 B0, set xnack (retry) on as default */
1020 	if (gc_ver == IP_VERSION(12, 1, 0) && (adev->rev_id & 0xf) == 0x1)
1021 		noretry_default = false;
1022 	if (!amdgpu_sriov_xnack_support(adev))
1023 		gmc->noretry = 1;
1024 	else
1025 		gmc->noretry = (amdgpu_noretry == -1) ? noretry_default : amdgpu_noretry;
1026 }
1027 
1028 void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
1029 				   bool enable)
1030 {
1031 	struct amdgpu_vmhub *hub;
1032 	u32 tmp, reg, i;
1033 
1034 	hub = &adev->vmhub[hub_type];
1035 	for (i = 0; i < 16; i++) {
1036 		reg = hub->vm_context0_cntl + hub->ctx_distance * i;
1037 
1038 		tmp = (hub_type == AMDGPU_GFXHUB(0)) ?
1039 			RREG32_SOC15_IP(GC, reg) :
1040 			RREG32_SOC15_IP(MMHUB, reg);
1041 
1042 		if (enable)
1043 			tmp |= hub->vm_cntx_cntl_vm_fault;
1044 		else
1045 			tmp &= ~hub->vm_cntx_cntl_vm_fault;
1046 
1047 		(hub_type == AMDGPU_GFXHUB(0)) ?
1048 			WREG32_SOC15_IP(GC, reg, tmp) :
1049 			WREG32_SOC15_IP(MMHUB, reg, tmp);
1050 	}
1051 }
1052 
1053 void amdgpu_gmc_init_vga_resv_regions(struct amdgpu_device *adev)
1054 {
1055 	unsigned size;
1056 
1057 	if (adev->gmc.is_app_apu)
1058 		return;
1059 
1060 	/*
1061 	 * Some ASICs need to reserve a region of video memory to avoid access
1062 	 * from driver
1063 	 */
1064 	/*
1065 	 * TODO:
1066 	 * Currently there is a bug where some memory client outside
1067 	 * of the driver writes to first 8M of VRAM on S3 resume,
1068 	 * this overrides GART which by default gets placed in first 8M and
1069 	 * causes VM_FAULTS once GTT is accessed.
1070 	 * Keep the stolen memory reservation until the while this is not solved.
1071 	 */
1072 	switch (adev->asic_type) {
1073 	case CHIP_VEGA10:
1074 		adev->mman.keep_stolen_vga_memory = true;
1075 		/*
1076 		 * VEGA10 SRIOV VF with MS_HYPERV host needs some firmware reserved area.
1077 		 */
1078 #ifdef CONFIG_X86
1079 		if (amdgpu_sriov_vf(adev) && hypervisor_is_type(X86_HYPER_MS_HYPERV)) {
1080 			amdgpu_ttm_init_vram_resv(adev, AMDGPU_RESV_STOLEN_RESERVED,
1081 						  0x500000, 0x200000, false);
1082 		}
1083 #endif
1084 		break;
1085 	case CHIP_RAVEN:
1086 	case CHIP_RENOIR:
1087 		adev->mman.keep_stolen_vga_memory = true;
1088 		break;
1089 	case CHIP_POLARIS10:
1090 	case CHIP_POLARIS11:
1091 	case CHIP_POLARIS12:
1092 		/* MacBookPros with switchable graphics put VRAM at 0 when
1093 		 * the iGPU is enabled which results in cursor issues if
1094 		 * the cursor ends up at 0.  Reserve vram at 0 in that case.
1095 		 */
1096 		if (adev->gmc.vram_start == 0)
1097 			adev->mman.keep_stolen_vga_memory = true;
1098 		break;
1099 	default:
1100 		adev->mman.keep_stolen_vga_memory = false;
1101 		break;
1102 	}
1103 
1104 	if (amdgpu_sriov_vf(adev) ||
1105 	    !amdgpu_device_has_display_hardware(adev)) {
1106 		size = 0;
1107 	} else {
1108 		size = amdgpu_gmc_get_vbios_fb_size(adev);
1109 
1110 		if (adev->mman.keep_stolen_vga_memory)
1111 			size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
1112 	}
1113 
1114 	/* set to 0 if the pre-OS buffer uses up most of vram */
1115 	if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
1116 		size = 0;
1117 
1118 	if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
1119 		amdgpu_ttm_init_vram_resv(adev, AMDGPU_RESV_STOLEN_VGA,
1120 					  0, AMDGPU_VBIOS_VGA_ALLOCATION, false);
1121 		amdgpu_ttm_init_vram_resv(adev, AMDGPU_RESV_STOLEN_EXTENDED,
1122 					  AMDGPU_VBIOS_VGA_ALLOCATION,
1123 					  size - AMDGPU_VBIOS_VGA_ALLOCATION, false);
1124 	} else {
1125 		amdgpu_ttm_init_vram_resv(adev, AMDGPU_RESV_STOLEN_VGA,
1126 					  0, size, false);
1127 	}
1128 }
1129 
1130 /**
1131  * amdgpu_gmc_init_pdb0 - initialize PDB0
1132  *
1133  * @adev: amdgpu_device pointer
1134  *
1135  * This function is only used when GART page table is used
1136  * for FB address translatioin. In such a case, we construct
1137  * a 2-level system VM page table: PDB0->PTB, to cover both
1138  * VRAM of the hive and system memory.
1139  *
1140  * PDB0 is static, initialized once on driver initialization.
1141  * The first n entries of PDB0 are used as PTE by setting
1142  * P bit to 1, pointing to VRAM. The n+1'th entry points
1143  * to a big PTB covering system memory.
1144  *
1145  */
1146 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
1147 {
1148 	int i;
1149 	uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
1150 	/* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
1151 	 */
1152 	u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
1153 	u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
1154 	u64 vram_addr, vram_end;
1155 	u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
1156 	int idx;
1157 
1158 	if (!drm_dev_enter(adev_to_drm(adev), &idx))
1159 		return;
1160 
1161 	flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
1162 	flags |= AMDGPU_PTE_WRITEABLE;
1163 	flags |= AMDGPU_PTE_SNOOPED;
1164 	flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
1165 	flags |= AMDGPU_PDE_PTE_FLAG(adev);
1166 
1167 	vram_addr = adev->vm_manager.vram_base_offset;
1168 	if (!amdgpu_virt_xgmi_migrate_enabled(adev))
1169 		vram_addr -= adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1170 	vram_end = vram_addr + vram_size;
1171 
1172 	/* The first n PDE0 entries are used as PTE,
1173 	 * pointing to vram
1174 	 */
1175 	for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
1176 		amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
1177 
1178 	/* The n+1'th PDE0 entry points to a huge
1179 	 * PTB who has more than 512 entries each
1180 	 * pointing to a 4K system page
1181 	 */
1182 	flags = AMDGPU_PTE_VALID;
1183 	flags |= AMDGPU_PTE_SNOOPED | AMDGPU_PDE_BFS_FLAG(adev, 0);
1184 	/* Requires gart_ptb_gpu_pa to be 4K aligned */
1185 	amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);
1186 	drm_dev_exit(idx);
1187 }
1188 
1189 /**
1190  * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC
1191  * address
1192  *
1193  * @adev: amdgpu_device pointer
1194  * @mc_addr: MC address of buffer
1195  */
1196 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr)
1197 {
1198 	return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset;
1199 }
1200 
1201 /**
1202  * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from
1203  * GPU's view
1204  *
1205  * @adev: amdgpu_device pointer
1206  * @bo: amdgpu buffer object
1207  */
1208 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
1209 {
1210 	return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo));
1211 }
1212 
1213 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev)
1214 {
1215 	struct amdgpu_bo *vram_bo = NULL;
1216 	uint64_t vram_gpu = 0;
1217 	void *vram_ptr = NULL;
1218 
1219 	int ret, size = 0x100000;
1220 	uint8_t cptr[10];
1221 
1222 	ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
1223 				AMDGPU_GEM_DOMAIN_VRAM,
1224 				&vram_bo,
1225 				&vram_gpu,
1226 				&vram_ptr);
1227 	if (ret)
1228 		return ret;
1229 
1230 	memset(vram_ptr, 0x86, size);
1231 	memset(cptr, 0x86, 10);
1232 
1233 	/**
1234 	 * Check the start, the mid, and the end of the memory if the content of
1235 	 * each byte is the pattern "0x86". If yes, we suppose the vram bo is
1236 	 * workable.
1237 	 *
1238 	 * Note: If check the each byte of whole 1M bo, it will cost too many
1239 	 * seconds, so here, we just pick up three parts for emulation.
1240 	 */
1241 	ret = memcmp(vram_ptr, cptr, 10);
1242 	if (ret) {
1243 		ret = -EIO;
1244 		goto release_buffer;
1245 	}
1246 
1247 	ret = memcmp(vram_ptr + (size / 2), cptr, 10);
1248 	if (ret) {
1249 		ret = -EIO;
1250 		goto release_buffer;
1251 	}
1252 
1253 	ret = memcmp(vram_ptr + size - 10, cptr, 10);
1254 	if (ret) {
1255 		ret = -EIO;
1256 		goto release_buffer;
1257 	}
1258 
1259 release_buffer:
1260 	amdgpu_bo_free_kernel(&vram_bo, &vram_gpu,
1261 			&vram_ptr);
1262 
1263 	return ret;
1264 }
1265 
1266 static const char *nps_desc[] = {
1267 	[AMDGPU_NPS1_PARTITION_MODE] = "NPS1",
1268 	[AMDGPU_NPS2_PARTITION_MODE] = "NPS2",
1269 	[AMDGPU_NPS3_PARTITION_MODE] = "NPS3",
1270 	[AMDGPU_NPS4_PARTITION_MODE] = "NPS4",
1271 	[AMDGPU_NPS6_PARTITION_MODE] = "NPS6",
1272 	[AMDGPU_NPS8_PARTITION_MODE] = "NPS8",
1273 };
1274 
1275 static ssize_t available_memory_partition_show(struct device *dev,
1276 					       struct device_attribute *addr,
1277 					       char *buf)
1278 {
1279 	struct drm_device *ddev = dev_get_drvdata(dev);
1280 	struct amdgpu_device *adev = drm_to_adev(ddev);
1281 	int size = 0, mode;
1282 	char *sep = "";
1283 
1284 	for_each_inst(mode, adev->gmc.supported_nps_modes) {
1285 		size += sysfs_emit_at(buf, size, "%s%s", sep, nps_desc[mode]);
1286 		sep = ", ";
1287 	}
1288 	size += sysfs_emit_at(buf, size, "\n");
1289 
1290 	return size;
1291 }
1292 
1293 static ssize_t current_memory_partition_store(struct device *dev,
1294 					      struct device_attribute *attr,
1295 					      const char *buf, size_t count)
1296 {
1297 	struct drm_device *ddev = dev_get_drvdata(dev);
1298 	struct amdgpu_device *adev = drm_to_adev(ddev);
1299 	enum amdgpu_memory_partition mode;
1300 	struct amdgpu_hive_info *hive;
1301 	int i;
1302 
1303 	mode = UNKNOWN_MEMORY_PARTITION_MODE;
1304 	for_each_inst(i, adev->gmc.supported_nps_modes) {
1305 		if (!strncasecmp(nps_desc[i], buf, strlen(nps_desc[i]))) {
1306 			mode = i;
1307 			break;
1308 		}
1309 	}
1310 
1311 	if (mode == UNKNOWN_MEMORY_PARTITION_MODE)
1312 		return -EINVAL;
1313 
1314 	if (mode == adev->gmc.gmc_funcs->query_mem_partition_mode(adev)) {
1315 		dev_info(
1316 			adev->dev,
1317 			"requested NPS mode is same as current NPS mode, skipping\n");
1318 		return count;
1319 	}
1320 
1321 	/* If device is part of hive, all devices in the hive should request the
1322 	 * same mode. Hence store the requested mode in hive.
1323 	 */
1324 	hive = amdgpu_get_xgmi_hive(adev);
1325 	if (hive) {
1326 		atomic_set(&hive->requested_nps_mode, mode);
1327 		amdgpu_put_xgmi_hive(hive);
1328 	} else {
1329 		adev->gmc.requested_nps_mode = mode;
1330 	}
1331 
1332 	dev_info(
1333 		adev->dev,
1334 		"NPS mode change requested, please remove and reload the driver\n");
1335 
1336 	return count;
1337 }
1338 
1339 static ssize_t current_memory_partition_show(
1340 	struct device *dev, struct device_attribute *addr, char *buf)
1341 {
1342 	struct drm_device *ddev = dev_get_drvdata(dev);
1343 	struct amdgpu_device *adev = drm_to_adev(ddev);
1344 	enum amdgpu_memory_partition mode;
1345 
1346 	/* Only minimal precaution taken to reject requests while in reset */
1347 	if (amdgpu_in_reset(adev))
1348 		return -EPERM;
1349 
1350 	mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
1351 	if ((mode >= ARRAY_SIZE(nps_desc)) ||
1352 	    (BIT(mode) & AMDGPU_ALL_NPS_MASK) != BIT(mode))
1353 		return sysfs_emit(buf, "UNKNOWN\n");
1354 
1355 	return sysfs_emit(buf, "%s\n", nps_desc[mode]);
1356 }
1357 
1358 static DEVICE_ATTR_RW(current_memory_partition);
1359 static DEVICE_ATTR_RO(available_memory_partition);
1360 
1361 int amdgpu_gmc_sysfs_init(struct amdgpu_device *adev)
1362 {
1363 	bool nps_switch_support;
1364 	int r = 0;
1365 
1366 	if (!adev->gmc.gmc_funcs->query_mem_partition_mode)
1367 		return 0;
1368 
1369 	nps_switch_support = (hweight32(adev->gmc.supported_nps_modes &
1370 					AMDGPU_ALL_NPS_MASK) > 1);
1371 	if (!nps_switch_support)
1372 		dev_attr_current_memory_partition.attr.mode &=
1373 			~(S_IWUSR | S_IWGRP | S_IWOTH);
1374 	else
1375 		r = device_create_file(adev->dev,
1376 				       &dev_attr_available_memory_partition);
1377 
1378 	if (r)
1379 		return r;
1380 
1381 	return device_create_file(adev->dev,
1382 				  &dev_attr_current_memory_partition);
1383 }
1384 
1385 void amdgpu_gmc_sysfs_fini(struct amdgpu_device *adev)
1386 {
1387 	if (!adev->gmc.gmc_funcs->query_mem_partition_mode)
1388 		return;
1389 
1390 	device_remove_file(adev->dev, &dev_attr_current_memory_partition);
1391 	device_remove_file(adev->dev, &dev_attr_available_memory_partition);
1392 }
1393 
1394 int amdgpu_gmc_get_nps_memranges(struct amdgpu_device *adev,
1395 				 struct amdgpu_mem_partition_info *mem_ranges,
1396 				 uint8_t *exp_ranges)
1397 {
1398 	struct amdgpu_gmc_memrange ranges[AMDGPU_MAX_MEM_RANGES];
1399 	int range_cnt, ret, i, j;
1400 	uint32_t nps_type;
1401 	bool refresh;
1402 
1403 	if (!mem_ranges || !exp_ranges)
1404 		return -EINVAL;
1405 	range_cnt = AMDGPU_MAX_MEM_RANGES;
1406 	refresh = (adev->init_lvl->level != AMDGPU_INIT_LEVEL_MINIMAL_XGMI) &&
1407 		  (adev->gmc.reset_flags & AMDGPU_GMC_INIT_RESET_NPS);
1408 	ret = amdgpu_discovery_get_nps_info(adev, &nps_type, ranges, &range_cnt,
1409 					    refresh);
1410 
1411 	if (ret)
1412 		return ret;
1413 
1414 	/* TODO: For now, expect ranges and partition count to be the same.
1415 	 * Adjust if there are holes expected in any NPS domain.
1416 	 */
1417 	if (*exp_ranges && (range_cnt != *exp_ranges)) {
1418 		dev_warn(
1419 			adev->dev,
1420 			"NPS config mismatch - expected ranges: %d discovery - nps mode: %d, nps ranges: %d",
1421 			*exp_ranges, nps_type, range_cnt);
1422 		ret = -EINVAL;
1423 		goto err;
1424 	}
1425 
1426 	for (i = 0; i < range_cnt; ++i) {
1427 		if (ranges[i].base_address >= ranges[i].limit_address) {
1428 			dev_warn(
1429 				adev->dev,
1430 				"Invalid NPS range - nps mode: %d, range[%d]: base: %llx limit: %llx",
1431 				nps_type, i, ranges[i].base_address,
1432 				ranges[i].limit_address);
1433 			ret = -EINVAL;
1434 			goto err;
1435 		}
1436 
1437 		/* Check for overlaps, not expecting any now */
1438 		for (j = i - 1; j >= 0; j--) {
1439 			if (max(ranges[j].base_address,
1440 				ranges[i].base_address) <=
1441 			    min(ranges[j].limit_address,
1442 				ranges[i].limit_address)) {
1443 				dev_warn(
1444 					adev->dev,
1445 					"overlapping ranges detected [ %llx - %llx ] | [%llx - %llx]",
1446 					ranges[j].base_address,
1447 					ranges[j].limit_address,
1448 					ranges[i].base_address,
1449 					ranges[i].limit_address);
1450 				ret = -EINVAL;
1451 				goto err;
1452 			}
1453 		}
1454 
1455 		mem_ranges[i].range.fpfn =
1456 			(ranges[i].base_address -
1457 			 adev->vm_manager.vram_base_offset) >>
1458 			AMDGPU_GPU_PAGE_SHIFT;
1459 		mem_ranges[i].range.lpfn =
1460 			(ranges[i].limit_address -
1461 			 adev->vm_manager.vram_base_offset) >>
1462 			AMDGPU_GPU_PAGE_SHIFT;
1463 		mem_ranges[i].size =
1464 			ranges[i].limit_address - ranges[i].base_address + 1;
1465 	}
1466 
1467 	if (!*exp_ranges)
1468 		*exp_ranges = range_cnt;
1469 err:
1470 	return ret;
1471 }
1472 
1473 int amdgpu_gmc_request_memory_partition(struct amdgpu_device *adev,
1474 					int nps_mode)
1475 {
1476 	/* Not supported on VF devices and APUs */
1477 	if (amdgpu_sriov_vf(adev) || (adev->flags & AMD_IS_APU))
1478 		return -EOPNOTSUPP;
1479 
1480 	if (!adev->psp.funcs) {
1481 		dev_err(adev->dev,
1482 			"PSP interface not available for nps mode change request");
1483 		return -EINVAL;
1484 	}
1485 
1486 	return psp_memory_partition(&adev->psp, nps_mode);
1487 }
1488 
1489 static inline bool amdgpu_gmc_need_nps_switch_req(struct amdgpu_device *adev,
1490 						  int req_nps_mode,
1491 						  int cur_nps_mode)
1492 {
1493 	return (((BIT(req_nps_mode) & adev->gmc.supported_nps_modes) ==
1494 			BIT(req_nps_mode)) &&
1495 		req_nps_mode != cur_nps_mode);
1496 }
1497 
1498 void amdgpu_gmc_prepare_nps_mode_change(struct amdgpu_device *adev)
1499 {
1500 	int req_nps_mode, cur_nps_mode, r;
1501 	struct amdgpu_hive_info *hive;
1502 
1503 	if (amdgpu_sriov_vf(adev) || !adev->gmc.supported_nps_modes ||
1504 	    !adev->gmc.gmc_funcs->request_mem_partition_mode)
1505 		return;
1506 
1507 	cur_nps_mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
1508 	hive = amdgpu_get_xgmi_hive(adev);
1509 	if (hive) {
1510 		req_nps_mode = atomic_read(&hive->requested_nps_mode);
1511 		if (!amdgpu_gmc_need_nps_switch_req(adev, req_nps_mode,
1512 						    cur_nps_mode)) {
1513 			amdgpu_put_xgmi_hive(hive);
1514 			return;
1515 		}
1516 		r = amdgpu_xgmi_request_nps_change(adev, hive, req_nps_mode);
1517 		amdgpu_put_xgmi_hive(hive);
1518 		goto out;
1519 	}
1520 
1521 	req_nps_mode = adev->gmc.requested_nps_mode;
1522 	if (!amdgpu_gmc_need_nps_switch_req(adev, req_nps_mode, cur_nps_mode))
1523 		return;
1524 
1525 	/* even if this fails, we should let driver unload w/o blocking */
1526 	r = adev->gmc.gmc_funcs->request_mem_partition_mode(adev, req_nps_mode);
1527 out:
1528 	if (r)
1529 		dev_err(adev->dev, "NPS mode change request failed\n");
1530 	else
1531 		dev_info(
1532 			adev->dev,
1533 			"NPS mode change request done, reload driver to complete the change\n");
1534 }
1535 
1536 bool amdgpu_gmc_need_reset_on_init(struct amdgpu_device *adev)
1537 {
1538 	if (adev->gmc.gmc_funcs->need_reset_on_init)
1539 		return adev->gmc.gmc_funcs->need_reset_on_init(adev);
1540 
1541 	return false;
1542 }
1543 
1544 enum amdgpu_memory_partition
1545 amdgpu_gmc_get_vf_memory_partition(struct amdgpu_device *adev)
1546 {
1547 	switch (adev->gmc.num_mem_partitions) {
1548 	case 0:
1549 		return UNKNOWN_MEMORY_PARTITION_MODE;
1550 	case 1:
1551 		return AMDGPU_NPS1_PARTITION_MODE;
1552 	case 2:
1553 		return AMDGPU_NPS2_PARTITION_MODE;
1554 	case 4:
1555 		return AMDGPU_NPS4_PARTITION_MODE;
1556 	case 8:
1557 		return AMDGPU_NPS8_PARTITION_MODE;
1558 	default:
1559 		return AMDGPU_NPS1_PARTITION_MODE;
1560 	}
1561 }
1562 
1563 enum amdgpu_memory_partition
1564 amdgpu_gmc_get_memory_partition(struct amdgpu_device *adev, u32 *supp_modes)
1565 {
1566 	enum amdgpu_memory_partition mode = UNKNOWN_MEMORY_PARTITION_MODE;
1567 
1568 	if (adev->nbio.funcs &&
1569 	    adev->nbio.funcs->get_memory_partition_mode)
1570 		mode = adev->nbio.funcs->get_memory_partition_mode(adev,
1571 								   supp_modes);
1572 	else
1573 		dev_warn(adev->dev, "memory partition mode query is not supported\n");
1574 
1575 	return mode;
1576 }
1577 
1578 enum amdgpu_memory_partition
1579 amdgpu_gmc_query_memory_partition(struct amdgpu_device *adev)
1580 {
1581 	if (amdgpu_sriov_vf(adev))
1582 		return amdgpu_gmc_get_vf_memory_partition(adev);
1583 	else
1584 		return amdgpu_gmc_get_memory_partition(adev, NULL);
1585 }
1586 
1587 static bool amdgpu_gmc_validate_partition_info(struct amdgpu_device *adev)
1588 {
1589 	enum amdgpu_memory_partition mode;
1590 	u32 supp_modes;
1591 	bool valid;
1592 
1593 	mode = amdgpu_gmc_get_memory_partition(adev, &supp_modes);
1594 
1595 	/* Mode detected by hardware not present in supported modes */
1596 	if ((mode != UNKNOWN_MEMORY_PARTITION_MODE) &&
1597 	    !(BIT(mode - 1) & supp_modes))
1598 		return false;
1599 
1600 	switch (mode) {
1601 	case UNKNOWN_MEMORY_PARTITION_MODE:
1602 	case AMDGPU_NPS1_PARTITION_MODE:
1603 		valid = (adev->gmc.num_mem_partitions == 1);
1604 		break;
1605 	case AMDGPU_NPS2_PARTITION_MODE:
1606 		valid = (adev->gmc.num_mem_partitions == 2);
1607 		break;
1608 	case AMDGPU_NPS4_PARTITION_MODE:
1609 		valid = (adev->gmc.num_mem_partitions == 3 ||
1610 			 adev->gmc.num_mem_partitions == 4);
1611 		break;
1612 	case AMDGPU_NPS8_PARTITION_MODE:
1613 		valid = (adev->gmc.num_mem_partitions == 8);
1614 		break;
1615 	default:
1616 		valid = false;
1617 	}
1618 
1619 	return valid;
1620 }
1621 
1622 static bool amdgpu_gmc_is_node_present(int *node_ids, int num_ids, int nid)
1623 {
1624 	int i;
1625 
1626 	/* Check if node with id 'nid' is present in 'node_ids' array */
1627 	for (i = 0; i < num_ids; ++i)
1628 		if (node_ids[i] == nid)
1629 			return true;
1630 
1631 	return false;
1632 }
1633 
1634 static void
1635 amdgpu_gmc_init_acpi_mem_ranges(struct amdgpu_device *adev,
1636 				struct amdgpu_mem_partition_info *mem_ranges)
1637 {
1638 	struct amdgpu_numa_info numa_info;
1639 	int node_ids[AMDGPU_MAX_MEM_RANGES];
1640 	int num_ranges = 0, ret;
1641 	int num_xcc, xcc_id;
1642 	uint32_t xcc_mask;
1643 
1644 	num_xcc = NUM_XCC(adev->gfx.xcc_mask);
1645 	xcc_mask = (1U << num_xcc) - 1;
1646 
1647 	for_each_inst(xcc_id, xcc_mask)	{
1648 		ret = amdgpu_acpi_get_mem_info(adev, xcc_id, &numa_info);
1649 		if (ret)
1650 			continue;
1651 
1652 		if (numa_info.nid == NUMA_NO_NODE) {
1653 			mem_ranges[0].size = numa_info.size;
1654 			mem_ranges[0].numa.node = numa_info.nid;
1655 			num_ranges = 1;
1656 			break;
1657 		}
1658 
1659 		if (amdgpu_gmc_is_node_present(node_ids, num_ranges,
1660 					     numa_info.nid))
1661 			continue;
1662 
1663 		node_ids[num_ranges] = numa_info.nid;
1664 		mem_ranges[num_ranges].numa.node = numa_info.nid;
1665 		mem_ranges[num_ranges].size = numa_info.size;
1666 		++num_ranges;
1667 	}
1668 
1669 	adev->gmc.num_mem_partitions = num_ranges;
1670 }
1671 
1672 void amdgpu_gmc_init_sw_mem_ranges(struct amdgpu_device *adev,
1673 				   struct amdgpu_mem_partition_info *mem_ranges)
1674 {
1675 	enum amdgpu_memory_partition mode;
1676 	u32 start_addr = 0, size;
1677 	int i, r, l;
1678 
1679 	mode = amdgpu_gmc_query_memory_partition(adev);
1680 
1681 	switch (mode) {
1682 	case UNKNOWN_MEMORY_PARTITION_MODE:
1683 		adev->gmc.num_mem_partitions = 0;
1684 		break;
1685 	case AMDGPU_NPS1_PARTITION_MODE:
1686 		adev->gmc.num_mem_partitions = 1;
1687 		break;
1688 	case AMDGPU_NPS2_PARTITION_MODE:
1689 		adev->gmc.num_mem_partitions = 2;
1690 		break;
1691 	case AMDGPU_NPS4_PARTITION_MODE:
1692 		if (adev->flags & AMD_IS_APU)
1693 			adev->gmc.num_mem_partitions = 3;
1694 		else
1695 			adev->gmc.num_mem_partitions = 4;
1696 		break;
1697 	case AMDGPU_NPS8_PARTITION_MODE:
1698 		adev->gmc.num_mem_partitions = 8;
1699 		break;
1700 	default:
1701 		adev->gmc.num_mem_partitions = 1;
1702 		break;
1703 	}
1704 
1705 	/* Use NPS range info, if populated */
1706 	r = amdgpu_gmc_get_nps_memranges(adev, mem_ranges,
1707 					 &adev->gmc.num_mem_partitions);
1708 	if (!r) {
1709 		l = 0;
1710 		for (i = 1; i < adev->gmc.num_mem_partitions; ++i) {
1711 			if (mem_ranges[i].range.lpfn >
1712 			    mem_ranges[i - 1].range.lpfn)
1713 				l = i;
1714 		}
1715 
1716 	} else {
1717 		if (!adev->gmc.num_mem_partitions) {
1718 			dev_warn(adev->dev,
1719 				 "Not able to detect NPS mode, fall back to NPS1\n");
1720 			adev->gmc.num_mem_partitions = 1;
1721 		}
1722 		/* Fallback to sw based calculation */
1723 		size = (adev->gmc.real_vram_size + SZ_16M) >> AMDGPU_GPU_PAGE_SHIFT;
1724 		size /= adev->gmc.num_mem_partitions;
1725 
1726 		for (i = 0; i < adev->gmc.num_mem_partitions; ++i) {
1727 			mem_ranges[i].range.fpfn = start_addr;
1728 			mem_ranges[i].size =
1729 				((u64)size << AMDGPU_GPU_PAGE_SHIFT);
1730 			mem_ranges[i].range.lpfn = start_addr + size - 1;
1731 			start_addr += size;
1732 		}
1733 
1734 		l = adev->gmc.num_mem_partitions - 1;
1735 	}
1736 
1737 	/* Adjust the last one */
1738 	mem_ranges[l].range.lpfn =
1739 		(adev->gmc.real_vram_size >> AMDGPU_GPU_PAGE_SHIFT) - 1;
1740 	mem_ranges[l].size =
1741 		adev->gmc.real_vram_size -
1742 		((u64)mem_ranges[l].range.fpfn << AMDGPU_GPU_PAGE_SHIFT);
1743 }
1744 
1745 int amdgpu_gmc_init_mem_ranges(struct amdgpu_device *adev)
1746 {
1747 	bool valid;
1748 
1749 	adev->gmc.mem_partitions = kzalloc_objs(struct amdgpu_mem_partition_info,
1750 						AMDGPU_MAX_MEM_RANGES);
1751 	if (!adev->gmc.mem_partitions)
1752 		return -ENOMEM;
1753 
1754 	if (adev->gmc.is_app_apu)
1755 		amdgpu_gmc_init_acpi_mem_ranges(adev, adev->gmc.mem_partitions);
1756 	else
1757 		amdgpu_gmc_init_sw_mem_ranges(adev, adev->gmc.mem_partitions);
1758 
1759 	if (amdgpu_sriov_vf(adev))
1760 		valid = true;
1761 	else
1762 		valid = amdgpu_gmc_validate_partition_info(adev);
1763 	if (!valid) {
1764 		/* TODO: handle invalid case */
1765 		dev_warn(adev->dev,
1766 			 "Mem ranges not matching with hardware config\n");
1767 	}
1768 
1769 	return 0;
1770 }
1771 
1772 int amdgpu_gmc_get_vram_info(struct amdgpu_device *adev,
1773 		int *vram_width, int *vram_type, int *vram_vendor)
1774 {
1775 	int ret = 0;
1776 
1777 	if (adev->flags & AMD_IS_APU)
1778 		return amdgpu_atomfirmware_get_integrated_system_info(adev,
1779 							vram_width, vram_type, vram_vendor);
1780 	switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
1781 	case IP_VERSION(12, 0, 0):
1782 	case IP_VERSION(12, 0, 1):
1783 		return amdgpu_atomfirmware_get_umc_info(adev,
1784 								vram_width, vram_type, vram_vendor);
1785 	case IP_VERSION(9, 5, 0):
1786 	case IP_VERSION(9, 4, 4):
1787 	case IP_VERSION(9, 4, 3):
1788 		ret = amdgpu_atomfirmware_get_umc_info(adev,
1789 								vram_width, vram_type, vram_vendor);
1790 		if (vram_width && !ret)
1791 			*vram_width *= hweight32(adev->aid_mask);
1792 		return ret;
1793 	default:
1794 		return amdgpu_atomfirmware_get_vram_info(adev,
1795 								vram_width, vram_type, vram_vendor);
1796 	}
1797 	return 0;
1798 }
1799