Loading...
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
38#include <drm/drm_drv.h>
39#include <drm/ttm/ttm_tt.h>
40
41/**
42 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
43 *
44 * @adev: amdgpu_device pointer
45 *
46 * Allocate video memory for pdb0 and map it for CPU access
47 * Returns 0 for success, error for failure.
48 */
49int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
50{
51 int r;
52 struct amdgpu_bo_param bp;
53 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
54 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
55 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) - 1) >> pde0_page_shift;
56
57 memset(&bp, 0, sizeof(bp));
58 bp.size = PAGE_ALIGN((npdes + 1) * 8);
59 bp.byte_align = PAGE_SIZE;
60 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
61 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
62 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
63 bp.type = ttm_bo_type_kernel;
64 bp.resv = NULL;
65 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
66
67 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
68 if (r)
69 return r;
70
71 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
72 if (unlikely(r != 0))
73 goto bo_reserve_failure;
74
75 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
76 if (r)
77 goto bo_pin_failure;
78 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
79 if (r)
80 goto bo_kmap_failure;
81
82 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
83 return 0;
84
85bo_kmap_failure:
86 amdgpu_bo_unpin(adev->gmc.pdb0_bo);
87bo_pin_failure:
88 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
89bo_reserve_failure:
90 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
91 return r;
92}
93
94/**
95 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
96 *
97 * @bo: the BO to get the PDE for
98 * @level: the level in the PD hirarchy
99 * @addr: resulting addr
100 * @flags: resulting flags
101 *
102 * Get the address and flags to be used for a PDE (Page Directory Entry).
103 */
104void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
105 uint64_t *addr, uint64_t *flags)
106{
107 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
108
109 switch (bo->tbo.resource->mem_type) {
110 case TTM_PL_TT:
111 *addr = bo->tbo.ttm->dma_address[0];
112 break;
113 case TTM_PL_VRAM:
114 *addr = amdgpu_bo_gpu_offset(bo);
115 break;
116 default:
117 *addr = 0;
118 break;
119 }
120 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource);
121 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
122}
123
124/*
125 * amdgpu_gmc_pd_addr - return the address of the root directory
126 */
127uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
128{
129 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
130 uint64_t pd_addr;
131
132 /* TODO: move that into ASIC specific code */
133 if (adev->asic_type >= CHIP_VEGA10) {
134 uint64_t flags = AMDGPU_PTE_VALID;
135
136 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
137 pd_addr |= flags;
138 } else {
139 pd_addr = amdgpu_bo_gpu_offset(bo);
140 }
141 return pd_addr;
142}
143
144/**
145 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
146 *
147 * @adev: amdgpu_device pointer
148 * @cpu_pt_addr: cpu address of the page table
149 * @gpu_page_idx: entry in the page table to update
150 * @addr: dst addr to write into pte/pde
151 * @flags: access flags
152 *
153 * Update the page tables using CPU.
154 */
155int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
156 uint32_t gpu_page_idx, uint64_t addr,
157 uint64_t flags)
158{
159 void __iomem *ptr = (void *)cpu_pt_addr;
160 uint64_t value;
161
162 /*
163 * The following is for PTE only. GART does not have PDEs.
164 */
165 value = addr & 0x0000FFFFFFFFF000ULL;
166 value |= flags;
167 writeq(value, ptr + (gpu_page_idx * 8));
168
169 return 0;
170}
171
172/**
173 * amdgpu_gmc_agp_addr - return the address in the AGP address space
174 *
175 * @bo: TTM BO which needs the address, must be in GTT domain
176 *
177 * Tries to figure out how to access the BO through the AGP aperture. Returns
178 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
179 */
180uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
181{
182 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
183
184 if (!bo->ttm)
185 return AMDGPU_BO_INVALID_OFFSET;
186
187 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
188 return AMDGPU_BO_INVALID_OFFSET;
189
190 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
191 return AMDGPU_BO_INVALID_OFFSET;
192
193 return adev->gmc.agp_start + bo->ttm->dma_address[0];
194}
195
196/**
197 * amdgpu_gmc_vram_location - try to find VRAM location
198 *
199 * @adev: amdgpu device structure holding all necessary information
200 * @mc: memory controller structure holding memory information
201 * @base: base address at which to put VRAM
202 *
203 * Function will try to place VRAM at base address provided
204 * as parameter.
205 */
206void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
207 u64 base)
208{
209 uint64_t vis_limit = (uint64_t)amdgpu_vis_vram_limit << 20;
210 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
211
212 mc->vram_start = base;
213 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
214 if (limit < mc->real_vram_size)
215 mc->real_vram_size = limit;
216
217 if (vis_limit && vis_limit < mc->visible_vram_size)
218 mc->visible_vram_size = vis_limit;
219
220 if (mc->real_vram_size < mc->visible_vram_size)
221 mc->visible_vram_size = mc->real_vram_size;
222
223 if (mc->xgmi.num_physical_nodes == 0) {
224 mc->fb_start = mc->vram_start;
225 mc->fb_end = mc->vram_end;
226 }
227 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
228 mc->mc_vram_size >> 20, mc->vram_start,
229 mc->vram_end, mc->real_vram_size >> 20);
230}
231
232/** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
233 *
234 * @adev: amdgpu device structure holding all necessary information
235 * @mc: memory controller structure holding memory information
236 *
237 * This function is only used if use GART for FB translation. In such
238 * case, we use sysvm aperture (vmid0 page tables) for both vram
239 * and gart (aka system memory) access.
240 *
241 * GPUVM (and our organization of vmid0 page tables) require sysvm
242 * aperture to be placed at a location aligned with 8 times of native
243 * page size. For example, if vm_context0_cntl.page_table_block_size
244 * is 12, then native page size is 8G (2M*2^12), sysvm should start
245 * with a 64G aligned address. For simplicity, we just put sysvm at
246 * address 0. So vram start at address 0 and gart is right after vram.
247 */
248void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
249{
250 u64 hive_vram_start = 0;
251 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
252 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
253 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
254 mc->gart_start = hive_vram_end + 1;
255 mc->gart_end = mc->gart_start + mc->gart_size - 1;
256 mc->fb_start = hive_vram_start;
257 mc->fb_end = hive_vram_end;
258 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
259 mc->mc_vram_size >> 20, mc->vram_start,
260 mc->vram_end, mc->real_vram_size >> 20);
261 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
262 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
263}
264
265/**
266 * amdgpu_gmc_gart_location - try to find GART location
267 *
268 * @adev: amdgpu device structure holding all necessary information
269 * @mc: memory controller structure holding memory information
270 * @gart_placement: GART placement policy with respect to VRAM
271 *
272 * Function will place try to place GART before or after VRAM.
273 * If GART size is bigger than space left then we ajust GART size.
274 * Thus function will never fails.
275 */
276void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
277 enum amdgpu_gart_placement gart_placement)
278{
279 const uint64_t four_gb = 0x100000000ULL;
280 u64 size_af, size_bf;
281 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
282 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
283
284 /* VCE doesn't like it when BOs cross a 4GB segment, so align
285 * the GART base on a 4GB boundary as well.
286 */
287 size_bf = mc->fb_start;
288 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
289
290 if (mc->gart_size > max(size_bf, size_af)) {
291 dev_warn(adev->dev, "limiting GART\n");
292 mc->gart_size = max(size_bf, size_af);
293 }
294
295 switch (gart_placement) {
296 case AMDGPU_GART_PLACEMENT_HIGH:
297 mc->gart_start = max_mc_address - mc->gart_size + 1;
298 break;
299 case AMDGPU_GART_PLACEMENT_LOW:
300 mc->gart_start = 0;
301 break;
302 case AMDGPU_GART_PLACEMENT_BEST_FIT:
303 default:
304 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
305 (size_af < mc->gart_size))
306 mc->gart_start = 0;
307 else
308 mc->gart_start = max_mc_address - mc->gart_size + 1;
309 break;
310 }
311
312 mc->gart_start &= ~(four_gb - 1);
313 mc->gart_end = mc->gart_start + mc->gart_size - 1;
314 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
315 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
316}
317
318/**
319 * amdgpu_gmc_agp_location - try to find AGP location
320 * @adev: amdgpu device structure holding all necessary information
321 * @mc: memory controller structure holding memory information
322 *
323 * Function will place try to find a place for the AGP BAR in the MC address
324 * space.
325 *
326 * AGP BAR will be assigned the largest available hole in the address space.
327 * Should be called after VRAM and GART locations are setup.
328 */
329void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
330{
331 const uint64_t sixteen_gb = 1ULL << 34;
332 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
333 u64 size_af, size_bf;
334
335 if (mc->fb_start > mc->gart_start) {
336 size_bf = (mc->fb_start & sixteen_gb_mask) -
337 ALIGN(mc->gart_end + 1, sixteen_gb);
338 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
339 } else {
340 size_bf = mc->fb_start & sixteen_gb_mask;
341 size_af = (mc->gart_start & sixteen_gb_mask) -
342 ALIGN(mc->fb_end + 1, sixteen_gb);
343 }
344
345 if (size_bf > size_af) {
346 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
347 mc->agp_size = size_bf;
348 } else {
349 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
350 mc->agp_size = size_af;
351 }
352
353 mc->agp_end = mc->agp_start + mc->agp_size - 1;
354 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
355 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
356}
357
358/**
359 * amdgpu_gmc_set_agp_default - Set the default AGP aperture value.
360 * @adev: amdgpu device structure holding all necessary information
361 * @mc: memory controller structure holding memory information
362 *
363 * To disable the AGP aperture, you need to set the start to a larger
364 * value than the end. This function sets the default value which
365 * can then be overridden using amdgpu_gmc_agp_location() if you want
366 * to enable the AGP aperture on a specific chip.
367 *
368 */
369void amdgpu_gmc_set_agp_default(struct amdgpu_device *adev,
370 struct amdgpu_gmc *mc)
371{
372 mc->agp_start = 0xffffffffffff;
373 mc->agp_end = 0;
374 mc->agp_size = 0;
375}
376
377/**
378 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid
379 *
380 * @addr: 48 bit physical address, page aligned (36 significant bits)
381 * @pasid: 16 bit process address space identifier
382 */
383static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid)
384{
385 return addr << 4 | pasid;
386}
387
388/**
389 * amdgpu_gmc_filter_faults - filter VM faults
390 *
391 * @adev: amdgpu device structure
392 * @ih: interrupt ring that the fault received from
393 * @addr: address of the VM fault
394 * @pasid: PASID of the process causing the fault
395 * @timestamp: timestamp of the fault
396 *
397 * Returns:
398 * True if the fault was filtered and should not be processed further.
399 * False if the fault is a new one and needs to be handled.
400 */
401bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
402 struct amdgpu_ih_ring *ih, uint64_t addr,
403 uint16_t pasid, uint64_t timestamp)
404{
405 struct amdgpu_gmc *gmc = &adev->gmc;
406 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid);
407 struct amdgpu_gmc_fault *fault;
408 uint32_t hash;
409
410 /* Stale retry fault if timestamp goes backward */
411 if (amdgpu_ih_ts_after(timestamp, ih->processed_timestamp))
412 return true;
413
414 /* If we don't have space left in the ring buffer return immediately */
415 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
416 AMDGPU_GMC_FAULT_TIMEOUT;
417 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
418 return true;
419
420 /* Try to find the fault in the hash */
421 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
422 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
423 while (fault->timestamp >= stamp) {
424 uint64_t tmp;
425
426 if (atomic64_read(&fault->key) == key) {
427 /*
428 * if we get a fault which is already present in
429 * the fault_ring and the timestamp of
430 * the fault is after the expired timestamp,
431 * then this is a new fault that needs to be added
432 * into the fault ring.
433 */
434 if (fault->timestamp_expiry != 0 &&
435 amdgpu_ih_ts_after(fault->timestamp_expiry,
436 timestamp))
437 break;
438 else
439 return true;
440 }
441
442 tmp = fault->timestamp;
443 fault = &gmc->fault_ring[fault->next];
444
445 /* Check if the entry was reused */
446 if (fault->timestamp >= tmp)
447 break;
448 }
449
450 /* Add the fault to the ring */
451 fault = &gmc->fault_ring[gmc->last_fault];
452 atomic64_set(&fault->key, key);
453 fault->timestamp = timestamp;
454
455 /* And update the hash */
456 fault->next = gmc->fault_hash[hash].idx;
457 gmc->fault_hash[hash].idx = gmc->last_fault++;
458 return false;
459}
460
461/**
462 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter
463 *
464 * @adev: amdgpu device structure
465 * @addr: address of the VM fault
466 * @pasid: PASID of the process causing the fault
467 *
468 * Remove the address from fault filter, then future vm fault on this address
469 * will pass to retry fault handler to recover.
470 */
471void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
472 uint16_t pasid)
473{
474 struct amdgpu_gmc *gmc = &adev->gmc;
475 uint64_t key = amdgpu_gmc_fault_key(addr, pasid);
476 struct amdgpu_ih_ring *ih;
477 struct amdgpu_gmc_fault *fault;
478 uint32_t last_wptr;
479 uint64_t last_ts;
480 uint32_t hash;
481 uint64_t tmp;
482
483 if (adev->irq.retry_cam_enabled)
484 return;
485
486 ih = &adev->irq.ih1;
487 /* Get the WPTR of the last entry in IH ring */
488 last_wptr = amdgpu_ih_get_wptr(adev, ih);
489 /* Order wptr with ring data. */
490 rmb();
491 /* Get the timetamp of the last entry in IH ring */
492 last_ts = amdgpu_ih_decode_iv_ts(adev, ih, last_wptr, -1);
493
494 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
495 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
496 do {
497 if (atomic64_read(&fault->key) == key) {
498 /*
499 * Update the timestamp when this fault
500 * expired.
501 */
502 fault->timestamp_expiry = last_ts;
503 break;
504 }
505
506 tmp = fault->timestamp;
507 fault = &gmc->fault_ring[fault->next];
508 } while (fault->timestamp < tmp);
509}
510
511int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev)
512{
513 int r;
514
515 /* umc ras block */
516 r = amdgpu_umc_ras_sw_init(adev);
517 if (r)
518 return r;
519
520 /* mmhub ras block */
521 r = amdgpu_mmhub_ras_sw_init(adev);
522 if (r)
523 return r;
524
525 /* hdp ras block */
526 r = amdgpu_hdp_ras_sw_init(adev);
527 if (r)
528 return r;
529
530 /* mca.x ras block */
531 r = amdgpu_mca_mp0_ras_sw_init(adev);
532 if (r)
533 return r;
534
535 r = amdgpu_mca_mp1_ras_sw_init(adev);
536 if (r)
537 return r;
538
539 r = amdgpu_mca_mpio_ras_sw_init(adev);
540 if (r)
541 return r;
542
543 /* xgmi ras block */
544 r = amdgpu_xgmi_ras_sw_init(adev);
545 if (r)
546 return r;
547
548 return 0;
549}
550
551int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
552{
553 return 0;
554}
555
556void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
557{
558
559}
560
561 /*
562 * The latest engine allocation on gfx9/10 is:
563 * Engine 2, 3: firmware
564 * Engine 0, 1, 4~16: amdgpu ring,
565 * subject to change when ring number changes
566 * Engine 17: Gart flushes
567 */
568#define AMDGPU_VMHUB_INV_ENG_BITMAP 0x1FFF3
569
570int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
571{
572 struct amdgpu_ring *ring;
573 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] = {0};
574 unsigned i;
575 unsigned vmhub, inv_eng;
576
577 /* init the vm inv eng for all vmhubs */
578 for_each_set_bit(i, adev->vmhubs_mask, AMDGPU_MAX_VMHUBS) {
579 vm_inv_engs[i] = AMDGPU_VMHUB_INV_ENG_BITMAP;
580 /* reserve engine 5 for firmware */
581 if (adev->enable_mes)
582 vm_inv_engs[i] &= ~(1 << 5);
583 /* reserve mmhub engine 3 for firmware */
584 if (adev->enable_umsch_mm)
585 vm_inv_engs[i] &= ~(1 << 3);
586 }
587
588 for (i = 0; i < adev->num_rings; ++i) {
589 ring = adev->rings[i];
590 vmhub = ring->vm_hub;
591
592 if (ring == &adev->mes.ring ||
593 ring == &adev->umsch_mm.ring)
594 continue;
595
596 inv_eng = ffs(vm_inv_engs[vmhub]);
597 if (!inv_eng) {
598 dev_err(adev->dev, "no VM inv eng for ring %s\n",
599 ring->name);
600 return -EINVAL;
601 }
602
603 ring->vm_inv_eng = inv_eng - 1;
604 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
605
606 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
607 ring->name, ring->vm_inv_eng, ring->vm_hub);
608 }
609
610 return 0;
611}
612
613void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
614 uint32_t vmhub, uint32_t flush_type)
615{
616 struct amdgpu_ring *ring = adev->mman.buffer_funcs_ring;
617 struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
618 struct dma_fence *fence;
619 struct amdgpu_job *job;
620 int r;
621
622 if (!hub->sdma_invalidation_workaround || vmid ||
623 !adev->mman.buffer_funcs_enabled ||
624 !adev->ib_pool_ready || amdgpu_in_reset(adev) ||
625 !ring->sched.ready) {
626
627 /*
628 * A GPU reset should flush all TLBs anyway, so no need to do
629 * this while one is ongoing.
630 */
631 if (!down_read_trylock(&adev->reset_domain->sem))
632 return;
633
634 if (adev->gmc.flush_tlb_needs_extra_type_2)
635 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid,
636 vmhub, 2);
637
638 if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2)
639 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid,
640 vmhub, 0);
641
642 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, vmid, vmhub,
643 flush_type);
644 up_read(&adev->reset_domain->sem);
645 return;
646 }
647
648 /* The SDMA on Navi 1x has a bug which can theoretically result in memory
649 * corruption if an invalidation happens at the same time as an VA
650 * translation. Avoid this by doing the invalidation from the SDMA
651 * itself at least for GART.
652 */
653 mutex_lock(&adev->mman.gtt_window_lock);
654 r = amdgpu_job_alloc_with_ib(ring->adev, &adev->mman.high_pr,
655 AMDGPU_FENCE_OWNER_UNDEFINED,
656 16 * 4, AMDGPU_IB_POOL_IMMEDIATE,
657 &job);
658 if (r)
659 goto error_alloc;
660
661 job->vm_pd_addr = amdgpu_gmc_pd_addr(adev->gart.bo);
662 job->vm_needs_flush = true;
663 job->ibs->ptr[job->ibs->length_dw++] = ring->funcs->nop;
664 amdgpu_ring_pad_ib(ring, &job->ibs[0]);
665 fence = amdgpu_job_submit(job);
666 mutex_unlock(&adev->mman.gtt_window_lock);
667
668 dma_fence_wait(fence, false);
669 dma_fence_put(fence);
670
671 return;
672
673error_alloc:
674 mutex_unlock(&adev->mman.gtt_window_lock);
675 dev_err(adev->dev, "Error flushing GPU TLB using the SDMA (%d)!\n", r);
676}
677
678int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid,
679 uint32_t flush_type, bool all_hub,
680 uint32_t inst)
681{
682 u32 usec_timeout = amdgpu_sriov_vf(adev) ? SRIOV_USEC_TIMEOUT :
683 adev->usec_timeout;
684 struct amdgpu_ring *ring = &adev->gfx.kiq[inst].ring;
685 struct amdgpu_kiq *kiq = &adev->gfx.kiq[inst];
686 unsigned int ndw;
687 signed long r;
688 uint32_t seq;
689
690 if (!adev->gmc.flush_pasid_uses_kiq || !ring->sched.ready ||
691 !down_read_trylock(&adev->reset_domain->sem)) {
692
693 if (adev->gmc.flush_tlb_needs_extra_type_2)
694 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
695 2, all_hub,
696 inst);
697
698 if (adev->gmc.flush_tlb_needs_extra_type_0 && flush_type == 2)
699 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
700 0, all_hub,
701 inst);
702
703 adev->gmc.gmc_funcs->flush_gpu_tlb_pasid(adev, pasid,
704 flush_type, all_hub,
705 inst);
706 return 0;
707 }
708
709 /* 2 dwords flush + 8 dwords fence */
710 ndw = kiq->pmf->invalidate_tlbs_size + 8;
711
712 if (adev->gmc.flush_tlb_needs_extra_type_2)
713 ndw += kiq->pmf->invalidate_tlbs_size;
714
715 if (adev->gmc.flush_tlb_needs_extra_type_0)
716 ndw += kiq->pmf->invalidate_tlbs_size;
717
718 spin_lock(&adev->gfx.kiq[inst].ring_lock);
719 amdgpu_ring_alloc(ring, ndw);
720 if (adev->gmc.flush_tlb_needs_extra_type_2)
721 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 2, all_hub);
722
723 if (flush_type == 2 && adev->gmc.flush_tlb_needs_extra_type_0)
724 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, 0, all_hub);
725
726 kiq->pmf->kiq_invalidate_tlbs(ring, pasid, flush_type, all_hub);
727 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
728 if (r) {
729 amdgpu_ring_undo(ring);
730 spin_unlock(&adev->gfx.kiq[inst].ring_lock);
731 goto error_unlock_reset;
732 }
733
734 amdgpu_ring_commit(ring);
735 spin_unlock(&adev->gfx.kiq[inst].ring_lock);
736 r = amdgpu_fence_wait_polling(ring, seq, usec_timeout);
737 if (r < 1) {
738 dev_err(adev->dev, "wait for kiq fence error: %ld.\n", r);
739 r = -ETIME;
740 goto error_unlock_reset;
741 }
742 r = 0;
743
744error_unlock_reset:
745 up_read(&adev->reset_domain->sem);
746 return r;
747}
748
749void amdgpu_gmc_fw_reg_write_reg_wait(struct amdgpu_device *adev,
750 uint32_t reg0, uint32_t reg1,
751 uint32_t ref, uint32_t mask,
752 uint32_t xcc_inst)
753{
754 struct amdgpu_kiq *kiq = &adev->gfx.kiq[xcc_inst];
755 struct amdgpu_ring *ring = &kiq->ring;
756 signed long r, cnt = 0;
757 unsigned long flags;
758 uint32_t seq;
759
760 if (adev->mes.ring.sched.ready) {
761 amdgpu_mes_reg_write_reg_wait(adev, reg0, reg1,
762 ref, mask);
763 return;
764 }
765
766 spin_lock_irqsave(&kiq->ring_lock, flags);
767 amdgpu_ring_alloc(ring, 32);
768 amdgpu_ring_emit_reg_write_reg_wait(ring, reg0, reg1,
769 ref, mask);
770 r = amdgpu_fence_emit_polling(ring, &seq, MAX_KIQ_REG_WAIT);
771 if (r)
772 goto failed_undo;
773
774 amdgpu_ring_commit(ring);
775 spin_unlock_irqrestore(&kiq->ring_lock, flags);
776
777 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
778
779 /* don't wait anymore for IRQ context */
780 if (r < 1 && in_interrupt())
781 goto failed_kiq;
782
783 might_sleep();
784 while (r < 1 && cnt++ < MAX_KIQ_REG_TRY) {
785
786 msleep(MAX_KIQ_REG_BAILOUT_INTERVAL);
787 r = amdgpu_fence_wait_polling(ring, seq, MAX_KIQ_REG_WAIT);
788 }
789
790 if (cnt > MAX_KIQ_REG_TRY)
791 goto failed_kiq;
792
793 return;
794
795failed_undo:
796 amdgpu_ring_undo(ring);
797 spin_unlock_irqrestore(&kiq->ring_lock, flags);
798failed_kiq:
799 dev_err(adev->dev, "failed to write reg %x wait reg %x\n", reg0, reg1);
800}
801
802/**
803 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
804 * @adev: amdgpu_device pointer
805 *
806 * Check and set if an the device @adev supports Trusted Memory
807 * Zones (TMZ).
808 */
809void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
810{
811 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) {
812 /* RAVEN */
813 case IP_VERSION(9, 2, 2):
814 case IP_VERSION(9, 1, 0):
815 /* RENOIR looks like RAVEN */
816 case IP_VERSION(9, 3, 0):
817 /* GC 10.3.7 */
818 case IP_VERSION(10, 3, 7):
819 /* GC 11.0.1 */
820 case IP_VERSION(11, 0, 1):
821 if (amdgpu_tmz == 0) {
822 adev->gmc.tmz_enabled = false;
823 dev_info(adev->dev,
824 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
825 } else {
826 adev->gmc.tmz_enabled = true;
827 dev_info(adev->dev,
828 "Trusted Memory Zone (TMZ) feature enabled\n");
829 }
830 break;
831 case IP_VERSION(10, 1, 10):
832 case IP_VERSION(10, 1, 1):
833 case IP_VERSION(10, 1, 2):
834 case IP_VERSION(10, 1, 3):
835 case IP_VERSION(10, 3, 0):
836 case IP_VERSION(10, 3, 2):
837 case IP_VERSION(10, 3, 4):
838 case IP_VERSION(10, 3, 5):
839 case IP_VERSION(10, 3, 6):
840 /* VANGOGH */
841 case IP_VERSION(10, 3, 1):
842 /* YELLOW_CARP*/
843 case IP_VERSION(10, 3, 3):
844 case IP_VERSION(11, 0, 4):
845 case IP_VERSION(11, 5, 0):
846 case IP_VERSION(11, 5, 1):
847 /* Don't enable it by default yet.
848 */
849 if (amdgpu_tmz < 1) {
850 adev->gmc.tmz_enabled = false;
851 dev_info(adev->dev,
852 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
853 } else {
854 adev->gmc.tmz_enabled = true;
855 dev_info(adev->dev,
856 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
857 }
858 break;
859 default:
860 adev->gmc.tmz_enabled = false;
861 dev_info(adev->dev,
862 "Trusted Memory Zone (TMZ) feature not supported\n");
863 break;
864 }
865}
866
867/**
868 * amdgpu_gmc_noretry_set -- set per asic noretry defaults
869 * @adev: amdgpu_device pointer
870 *
871 * Set a per asic default for the no-retry parameter.
872 *
873 */
874void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
875{
876 struct amdgpu_gmc *gmc = &adev->gmc;
877 uint32_t gc_ver = amdgpu_ip_version(adev, GC_HWIP, 0);
878 bool noretry_default = (gc_ver == IP_VERSION(9, 0, 1) ||
879 gc_ver == IP_VERSION(9, 3, 0) ||
880 gc_ver == IP_VERSION(9, 4, 0) ||
881 gc_ver == IP_VERSION(9, 4, 1) ||
882 gc_ver == IP_VERSION(9, 4, 2) ||
883 gc_ver == IP_VERSION(9, 4, 3) ||
884 gc_ver >= IP_VERSION(10, 3, 0));
885
886 if (!amdgpu_sriov_xnack_support(adev))
887 gmc->noretry = 1;
888 else
889 gmc->noretry = (amdgpu_noretry == -1) ? noretry_default : amdgpu_noretry;
890}
891
892void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
893 bool enable)
894{
895 struct amdgpu_vmhub *hub;
896 u32 tmp, reg, i;
897
898 hub = &adev->vmhub[hub_type];
899 for (i = 0; i < 16; i++) {
900 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
901
902 tmp = (hub_type == AMDGPU_GFXHUB(0)) ?
903 RREG32_SOC15_IP(GC, reg) :
904 RREG32_SOC15_IP(MMHUB, reg);
905
906 if (enable)
907 tmp |= hub->vm_cntx_cntl_vm_fault;
908 else
909 tmp &= ~hub->vm_cntx_cntl_vm_fault;
910
911 (hub_type == AMDGPU_GFXHUB(0)) ?
912 WREG32_SOC15_IP(GC, reg, tmp) :
913 WREG32_SOC15_IP(MMHUB, reg, tmp);
914 }
915}
916
917void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
918{
919 unsigned size;
920
921 /*
922 * Some ASICs need to reserve a region of video memory to avoid access
923 * from driver
924 */
925 adev->mman.stolen_reserved_offset = 0;
926 adev->mman.stolen_reserved_size = 0;
927
928 /*
929 * TODO:
930 * Currently there is a bug where some memory client outside
931 * of the driver writes to first 8M of VRAM on S3 resume,
932 * this overrides GART which by default gets placed in first 8M and
933 * causes VM_FAULTS once GTT is accessed.
934 * Keep the stolen memory reservation until the while this is not solved.
935 */
936 switch (adev->asic_type) {
937 case CHIP_VEGA10:
938 adev->mman.keep_stolen_vga_memory = true;
939 /*
940 * VEGA10 SRIOV VF with MS_HYPERV host needs some firmware reserved area.
941 */
942#ifdef CONFIG_X86
943 if (amdgpu_sriov_vf(adev) && hypervisor_is_type(X86_HYPER_MS_HYPERV)) {
944 adev->mman.stolen_reserved_offset = 0x500000;
945 adev->mman.stolen_reserved_size = 0x200000;
946 }
947#endif
948 break;
949 case CHIP_RAVEN:
950 case CHIP_RENOIR:
951 adev->mman.keep_stolen_vga_memory = true;
952 break;
953 default:
954 adev->mman.keep_stolen_vga_memory = false;
955 break;
956 }
957
958 if (amdgpu_sriov_vf(adev) ||
959 !amdgpu_device_has_display_hardware(adev)) {
960 size = 0;
961 } else {
962 size = amdgpu_gmc_get_vbios_fb_size(adev);
963
964 if (adev->mman.keep_stolen_vga_memory)
965 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
966 }
967
968 /* set to 0 if the pre-OS buffer uses up most of vram */
969 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
970 size = 0;
971
972 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
973 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
974 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
975 } else {
976 adev->mman.stolen_vga_size = size;
977 adev->mman.stolen_extended_size = 0;
978 }
979}
980
981/**
982 * amdgpu_gmc_init_pdb0 - initialize PDB0
983 *
984 * @adev: amdgpu_device pointer
985 *
986 * This function is only used when GART page table is used
987 * for FB address translatioin. In such a case, we construct
988 * a 2-level system VM page table: PDB0->PTB, to cover both
989 * VRAM of the hive and system memory.
990 *
991 * PDB0 is static, initialized once on driver initialization.
992 * The first n entries of PDB0 are used as PTE by setting
993 * P bit to 1, pointing to VRAM. The n+1'th entry points
994 * to a big PTB covering system memory.
995 *
996 */
997void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
998{
999 int i;
1000 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
1001 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
1002 */
1003 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
1004 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
1005 u64 vram_addr = adev->vm_manager.vram_base_offset -
1006 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
1007 u64 vram_end = vram_addr + vram_size;
1008 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
1009 int idx;
1010
1011 if (!drm_dev_enter(adev_to_drm(adev), &idx))
1012 return;
1013
1014 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
1015 flags |= AMDGPU_PTE_WRITEABLE;
1016 flags |= AMDGPU_PTE_SNOOPED;
1017 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
1018 flags |= AMDGPU_PDE_PTE;
1019
1020 /* The first n PDE0 entries are used as PTE,
1021 * pointing to vram
1022 */
1023 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
1024 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
1025
1026 /* The n+1'th PDE0 entry points to a huge
1027 * PTB who has more than 512 entries each
1028 * pointing to a 4K system page
1029 */
1030 flags = AMDGPU_PTE_VALID;
1031 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED;
1032 /* Requires gart_ptb_gpu_pa to be 4K aligned */
1033 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);
1034 drm_dev_exit(idx);
1035}
1036
1037/**
1038 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC
1039 * address
1040 *
1041 * @adev: amdgpu_device pointer
1042 * @mc_addr: MC address of buffer
1043 */
1044uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr)
1045{
1046 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset;
1047}
1048
1049/**
1050 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from
1051 * GPU's view
1052 *
1053 * @adev: amdgpu_device pointer
1054 * @bo: amdgpu buffer object
1055 */
1056uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
1057{
1058 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo));
1059}
1060
1061/**
1062 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address
1063 * from CPU's view
1064 *
1065 * @adev: amdgpu_device pointer
1066 * @bo: amdgpu buffer object
1067 */
1068uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
1069{
1070 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base;
1071}
1072
1073int amdgpu_gmc_vram_checking(struct amdgpu_device *adev)
1074{
1075 struct amdgpu_bo *vram_bo = NULL;
1076 uint64_t vram_gpu = 0;
1077 void *vram_ptr = NULL;
1078
1079 int ret, size = 0x100000;
1080 uint8_t cptr[10];
1081
1082 ret = amdgpu_bo_create_kernel(adev, size, PAGE_SIZE,
1083 AMDGPU_GEM_DOMAIN_VRAM,
1084 &vram_bo,
1085 &vram_gpu,
1086 &vram_ptr);
1087 if (ret)
1088 return ret;
1089
1090 memset(vram_ptr, 0x86, size);
1091 memset(cptr, 0x86, 10);
1092
1093 /**
1094 * Check the start, the mid, and the end of the memory if the content of
1095 * each byte is the pattern "0x86". If yes, we suppose the vram bo is
1096 * workable.
1097 *
1098 * Note: If check the each byte of whole 1M bo, it will cost too many
1099 * seconds, so here, we just pick up three parts for emulation.
1100 */
1101 ret = memcmp(vram_ptr, cptr, 10);
1102 if (ret) {
1103 ret = -EIO;
1104 goto release_buffer;
1105 }
1106
1107 ret = memcmp(vram_ptr + (size / 2), cptr, 10);
1108 if (ret) {
1109 ret = -EIO;
1110 goto release_buffer;
1111 }
1112
1113 ret = memcmp(vram_ptr + size - 10, cptr, 10);
1114 if (ret) {
1115 ret = -EIO;
1116 goto release_buffer;
1117 }
1118
1119release_buffer:
1120 amdgpu_bo_free_kernel(&vram_bo, &vram_gpu,
1121 &vram_ptr);
1122
1123 return ret;
1124}
1125
1126static ssize_t current_memory_partition_show(
1127 struct device *dev, struct device_attribute *addr, char *buf)
1128{
1129 struct drm_device *ddev = dev_get_drvdata(dev);
1130 struct amdgpu_device *adev = drm_to_adev(ddev);
1131 enum amdgpu_memory_partition mode;
1132
1133 mode = adev->gmc.gmc_funcs->query_mem_partition_mode(adev);
1134 switch (mode) {
1135 case AMDGPU_NPS1_PARTITION_MODE:
1136 return sysfs_emit(buf, "NPS1\n");
1137 case AMDGPU_NPS2_PARTITION_MODE:
1138 return sysfs_emit(buf, "NPS2\n");
1139 case AMDGPU_NPS3_PARTITION_MODE:
1140 return sysfs_emit(buf, "NPS3\n");
1141 case AMDGPU_NPS4_PARTITION_MODE:
1142 return sysfs_emit(buf, "NPS4\n");
1143 case AMDGPU_NPS6_PARTITION_MODE:
1144 return sysfs_emit(buf, "NPS6\n");
1145 case AMDGPU_NPS8_PARTITION_MODE:
1146 return sysfs_emit(buf, "NPS8\n");
1147 default:
1148 return sysfs_emit(buf, "UNKNOWN\n");
1149 }
1150
1151 return sysfs_emit(buf, "UNKNOWN\n");
1152}
1153
1154static DEVICE_ATTR_RO(current_memory_partition);
1155
1156int amdgpu_gmc_sysfs_init(struct amdgpu_device *adev)
1157{
1158 if (!adev->gmc.gmc_funcs->query_mem_partition_mode)
1159 return 0;
1160
1161 return device_create_file(adev->dev,
1162 &dev_attr_current_memory_partition);
1163}
1164
1165void amdgpu_gmc_sysfs_fini(struct amdgpu_device *adev)
1166{
1167 device_remove_file(adev->dev, &dev_attr_current_memory_partition);
1168}
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
29#include "amdgpu.h"
30#include "amdgpu_gmc.h"
31#include "amdgpu_ras.h"
32#include "amdgpu_xgmi.h"
33
34#include <drm/drm_drv.h>
35
36/**
37 * amdgpu_gmc_pdb0_alloc - allocate vram for pdb0
38 *
39 * @adev: amdgpu_device pointer
40 *
41 * Allocate video memory for pdb0 and map it for CPU access
42 * Returns 0 for success, error for failure.
43 */
44int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev)
45{
46 int r;
47 struct amdgpu_bo_param bp;
48 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
49 uint32_t pde0_page_shift = adev->gmc.vmid0_page_table_block_size + 21;
50 uint32_t npdes = (vram_size + (1ULL << pde0_page_shift) -1) >> pde0_page_shift;
51
52 memset(&bp, 0, sizeof(bp));
53 bp.size = PAGE_ALIGN((npdes + 1) * 8);
54 bp.byte_align = PAGE_SIZE;
55 bp.domain = AMDGPU_GEM_DOMAIN_VRAM;
56 bp.flags = AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED |
57 AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS;
58 bp.type = ttm_bo_type_kernel;
59 bp.resv = NULL;
60 bp.bo_ptr_size = sizeof(struct amdgpu_bo);
61
62 r = amdgpu_bo_create(adev, &bp, &adev->gmc.pdb0_bo);
63 if (r)
64 return r;
65
66 r = amdgpu_bo_reserve(adev->gmc.pdb0_bo, false);
67 if (unlikely(r != 0))
68 goto bo_reserve_failure;
69
70 r = amdgpu_bo_pin(adev->gmc.pdb0_bo, AMDGPU_GEM_DOMAIN_VRAM);
71 if (r)
72 goto bo_pin_failure;
73 r = amdgpu_bo_kmap(adev->gmc.pdb0_bo, &adev->gmc.ptr_pdb0);
74 if (r)
75 goto bo_kmap_failure;
76
77 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
78 return 0;
79
80bo_kmap_failure:
81 amdgpu_bo_unpin(adev->gmc.pdb0_bo);
82bo_pin_failure:
83 amdgpu_bo_unreserve(adev->gmc.pdb0_bo);
84bo_reserve_failure:
85 amdgpu_bo_unref(&adev->gmc.pdb0_bo);
86 return r;
87}
88
89/**
90 * amdgpu_gmc_get_pde_for_bo - get the PDE for a BO
91 *
92 * @bo: the BO to get the PDE for
93 * @level: the level in the PD hirarchy
94 * @addr: resulting addr
95 * @flags: resulting flags
96 *
97 * Get the address and flags to be used for a PDE (Page Directory Entry).
98 */
99void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
100 uint64_t *addr, uint64_t *flags)
101{
102 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
103
104 switch (bo->tbo.resource->mem_type) {
105 case TTM_PL_TT:
106 *addr = bo->tbo.ttm->dma_address[0];
107 break;
108 case TTM_PL_VRAM:
109 *addr = amdgpu_bo_gpu_offset(bo);
110 break;
111 default:
112 *addr = 0;
113 break;
114 }
115 *flags = amdgpu_ttm_tt_pde_flags(bo->tbo.ttm, bo->tbo.resource);
116 amdgpu_gmc_get_vm_pde(adev, level, addr, flags);
117}
118
119/*
120 * amdgpu_gmc_pd_addr - return the address of the root directory
121 */
122uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo)
123{
124 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
125 uint64_t pd_addr;
126
127 /* TODO: move that into ASIC specific code */
128 if (adev->asic_type >= CHIP_VEGA10) {
129 uint64_t flags = AMDGPU_PTE_VALID;
130
131 amdgpu_gmc_get_pde_for_bo(bo, -1, &pd_addr, &flags);
132 pd_addr |= flags;
133 } else {
134 pd_addr = amdgpu_bo_gpu_offset(bo);
135 }
136 return pd_addr;
137}
138
139/**
140 * amdgpu_gmc_set_pte_pde - update the page tables using CPU
141 *
142 * @adev: amdgpu_device pointer
143 * @cpu_pt_addr: cpu address of the page table
144 * @gpu_page_idx: entry in the page table to update
145 * @addr: dst addr to write into pte/pde
146 * @flags: access flags
147 *
148 * Update the page tables using CPU.
149 */
150int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
151 uint32_t gpu_page_idx, uint64_t addr,
152 uint64_t flags)
153{
154 void __iomem *ptr = (void *)cpu_pt_addr;
155 uint64_t value;
156 int idx;
157
158 if (!drm_dev_enter(&adev->ddev, &idx))
159 return 0;
160
161 /*
162 * The following is for PTE only. GART does not have PDEs.
163 */
164 value = addr & 0x0000FFFFFFFFF000ULL;
165 value |= flags;
166 writeq(value, ptr + (gpu_page_idx * 8));
167
168 drm_dev_exit(idx);
169
170 return 0;
171}
172
173/**
174 * amdgpu_gmc_agp_addr - return the address in the AGP address space
175 *
176 * @bo: TTM BO which needs the address, must be in GTT domain
177 *
178 * Tries to figure out how to access the BO through the AGP aperture. Returns
179 * AMDGPU_BO_INVALID_OFFSET if that is not possible.
180 */
181uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo)
182{
183 struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
184
185 if (bo->ttm->num_pages != 1 || bo->ttm->caching == ttm_cached)
186 return AMDGPU_BO_INVALID_OFFSET;
187
188 if (bo->ttm->dma_address[0] + PAGE_SIZE >= adev->gmc.agp_size)
189 return AMDGPU_BO_INVALID_OFFSET;
190
191 return adev->gmc.agp_start + bo->ttm->dma_address[0];
192}
193
194/**
195 * amdgpu_gmc_vram_location - try to find VRAM location
196 *
197 * @adev: amdgpu device structure holding all necessary information
198 * @mc: memory controller structure holding memory information
199 * @base: base address at which to put VRAM
200 *
201 * Function will try to place VRAM at base address provided
202 * as parameter.
203 */
204void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
205 u64 base)
206{
207 uint64_t limit = (uint64_t)amdgpu_vram_limit << 20;
208
209 mc->vram_start = base;
210 mc->vram_end = mc->vram_start + mc->mc_vram_size - 1;
211 if (limit && limit < mc->real_vram_size)
212 mc->real_vram_size = limit;
213
214 if (mc->xgmi.num_physical_nodes == 0) {
215 mc->fb_start = mc->vram_start;
216 mc->fb_end = mc->vram_end;
217 }
218 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
219 mc->mc_vram_size >> 20, mc->vram_start,
220 mc->vram_end, mc->real_vram_size >> 20);
221}
222
223/** amdgpu_gmc_sysvm_location - place vram and gart in sysvm aperture
224 *
225 * @adev: amdgpu device structure holding all necessary information
226 * @mc: memory controller structure holding memory information
227 *
228 * This function is only used if use GART for FB translation. In such
229 * case, we use sysvm aperture (vmid0 page tables) for both vram
230 * and gart (aka system memory) access.
231 *
232 * GPUVM (and our organization of vmid0 page tables) require sysvm
233 * aperture to be placed at a location aligned with 8 times of native
234 * page size. For example, if vm_context0_cntl.page_table_block_size
235 * is 12, then native page size is 8G (2M*2^12), sysvm should start
236 * with a 64G aligned address. For simplicity, we just put sysvm at
237 * address 0. So vram start at address 0 and gart is right after vram.
238 */
239void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
240{
241 u64 hive_vram_start = 0;
242 u64 hive_vram_end = mc->xgmi.node_segment_size * mc->xgmi.num_physical_nodes - 1;
243 mc->vram_start = mc->xgmi.node_segment_size * mc->xgmi.physical_node_id;
244 mc->vram_end = mc->vram_start + mc->xgmi.node_segment_size - 1;
245 mc->gart_start = hive_vram_end + 1;
246 mc->gart_end = mc->gart_start + mc->gart_size - 1;
247 mc->fb_start = hive_vram_start;
248 mc->fb_end = hive_vram_end;
249 dev_info(adev->dev, "VRAM: %lluM 0x%016llX - 0x%016llX (%lluM used)\n",
250 mc->mc_vram_size >> 20, mc->vram_start,
251 mc->vram_end, mc->real_vram_size >> 20);
252 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
253 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
254}
255
256/**
257 * amdgpu_gmc_gart_location - try to find GART location
258 *
259 * @adev: amdgpu device structure holding all necessary information
260 * @mc: memory controller structure holding memory information
261 *
262 * Function will place try to place GART before or after VRAM.
263 * If GART size is bigger than space left then we ajust GART size.
264 * Thus function will never fails.
265 */
266void amdgpu_gmc_gart_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
267{
268 const uint64_t four_gb = 0x100000000ULL;
269 u64 size_af, size_bf;
270 /*To avoid the hole, limit the max mc address to AMDGPU_GMC_HOLE_START*/
271 u64 max_mc_address = min(adev->gmc.mc_mask, AMDGPU_GMC_HOLE_START - 1);
272
273 /* VCE doesn't like it when BOs cross a 4GB segment, so align
274 * the GART base on a 4GB boundary as well.
275 */
276 size_bf = mc->fb_start;
277 size_af = max_mc_address + 1 - ALIGN(mc->fb_end + 1, four_gb);
278
279 if (mc->gart_size > max(size_bf, size_af)) {
280 dev_warn(adev->dev, "limiting GART\n");
281 mc->gart_size = max(size_bf, size_af);
282 }
283
284 if ((size_bf >= mc->gart_size && size_bf < size_af) ||
285 (size_af < mc->gart_size))
286 mc->gart_start = 0;
287 else
288 mc->gart_start = max_mc_address - mc->gart_size + 1;
289
290 mc->gart_start &= ~(four_gb - 1);
291 mc->gart_end = mc->gart_start + mc->gart_size - 1;
292 dev_info(adev->dev, "GART: %lluM 0x%016llX - 0x%016llX\n",
293 mc->gart_size >> 20, mc->gart_start, mc->gart_end);
294}
295
296/**
297 * amdgpu_gmc_agp_location - try to find AGP location
298 * @adev: amdgpu device structure holding all necessary information
299 * @mc: memory controller structure holding memory information
300 *
301 * Function will place try to find a place for the AGP BAR in the MC address
302 * space.
303 *
304 * AGP BAR will be assigned the largest available hole in the address space.
305 * Should be called after VRAM and GART locations are setup.
306 */
307void amdgpu_gmc_agp_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc)
308{
309 const uint64_t sixteen_gb = 1ULL << 34;
310 const uint64_t sixteen_gb_mask = ~(sixteen_gb - 1);
311 u64 size_af, size_bf;
312
313 if (amdgpu_sriov_vf(adev)) {
314 mc->agp_start = 0xffffffffffff;
315 mc->agp_end = 0x0;
316 mc->agp_size = 0;
317
318 return;
319 }
320
321 if (mc->fb_start > mc->gart_start) {
322 size_bf = (mc->fb_start & sixteen_gb_mask) -
323 ALIGN(mc->gart_end + 1, sixteen_gb);
324 size_af = mc->mc_mask + 1 - ALIGN(mc->fb_end + 1, sixteen_gb);
325 } else {
326 size_bf = mc->fb_start & sixteen_gb_mask;
327 size_af = (mc->gart_start & sixteen_gb_mask) -
328 ALIGN(mc->fb_end + 1, sixteen_gb);
329 }
330
331 if (size_bf > size_af) {
332 mc->agp_start = (mc->fb_start - size_bf) & sixteen_gb_mask;
333 mc->agp_size = size_bf;
334 } else {
335 mc->agp_start = ALIGN(mc->fb_end + 1, sixteen_gb);
336 mc->agp_size = size_af;
337 }
338
339 mc->agp_end = mc->agp_start + mc->agp_size - 1;
340 dev_info(adev->dev, "AGP: %lluM 0x%016llX - 0x%016llX\n",
341 mc->agp_size >> 20, mc->agp_start, mc->agp_end);
342}
343
344/**
345 * amdgpu_gmc_fault_key - get hask key from vm fault address and pasid
346 *
347 * @addr: 48 bit physical address, page aligned (36 significant bits)
348 * @pasid: 16 bit process address space identifier
349 */
350static inline uint64_t amdgpu_gmc_fault_key(uint64_t addr, uint16_t pasid)
351{
352 return addr << 4 | pasid;
353}
354
355/**
356 * amdgpu_gmc_filter_faults - filter VM faults
357 *
358 * @adev: amdgpu device structure
359 * @addr: address of the VM fault
360 * @pasid: PASID of the process causing the fault
361 * @timestamp: timestamp of the fault
362 *
363 * Returns:
364 * True if the fault was filtered and should not be processed further.
365 * False if the fault is a new one and needs to be handled.
366 */
367bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr,
368 uint16_t pasid, uint64_t timestamp)
369{
370 struct amdgpu_gmc *gmc = &adev->gmc;
371 uint64_t stamp, key = amdgpu_gmc_fault_key(addr, pasid);
372 struct amdgpu_gmc_fault *fault;
373 uint32_t hash;
374
375 /* If we don't have space left in the ring buffer return immediately */
376 stamp = max(timestamp, AMDGPU_GMC_FAULT_TIMEOUT + 1) -
377 AMDGPU_GMC_FAULT_TIMEOUT;
378 if (gmc->fault_ring[gmc->last_fault].timestamp >= stamp)
379 return true;
380
381 /* Try to find the fault in the hash */
382 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
383 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
384 while (fault->timestamp >= stamp) {
385 uint64_t tmp;
386
387 if (atomic64_read(&fault->key) == key)
388 return true;
389
390 tmp = fault->timestamp;
391 fault = &gmc->fault_ring[fault->next];
392
393 /* Check if the entry was reused */
394 if (fault->timestamp >= tmp)
395 break;
396 }
397
398 /* Add the fault to the ring */
399 fault = &gmc->fault_ring[gmc->last_fault];
400 atomic64_set(&fault->key, key);
401 fault->timestamp = timestamp;
402
403 /* And update the hash */
404 fault->next = gmc->fault_hash[hash].idx;
405 gmc->fault_hash[hash].idx = gmc->last_fault++;
406 return false;
407}
408
409/**
410 * amdgpu_gmc_filter_faults_remove - remove address from VM faults filter
411 *
412 * @adev: amdgpu device structure
413 * @addr: address of the VM fault
414 * @pasid: PASID of the process causing the fault
415 *
416 * Remove the address from fault filter, then future vm fault on this address
417 * will pass to retry fault handler to recover.
418 */
419void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
420 uint16_t pasid)
421{
422 struct amdgpu_gmc *gmc = &adev->gmc;
423 uint64_t key = amdgpu_gmc_fault_key(addr, pasid);
424 struct amdgpu_gmc_fault *fault;
425 uint32_t hash;
426 uint64_t tmp;
427
428 hash = hash_64(key, AMDGPU_GMC_FAULT_HASH_ORDER);
429 fault = &gmc->fault_ring[gmc->fault_hash[hash].idx];
430 do {
431 if (atomic64_cmpxchg(&fault->key, key, 0) == key)
432 break;
433
434 tmp = fault->timestamp;
435 fault = &gmc->fault_ring[fault->next];
436 } while (fault->timestamp < tmp);
437}
438
439int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev)
440{
441 int r;
442
443 if (adev->umc.ras_funcs &&
444 adev->umc.ras_funcs->ras_late_init) {
445 r = adev->umc.ras_funcs->ras_late_init(adev);
446 if (r)
447 return r;
448 }
449
450 if (adev->mmhub.ras_funcs &&
451 adev->mmhub.ras_funcs->ras_late_init) {
452 r = adev->mmhub.ras_funcs->ras_late_init(adev);
453 if (r)
454 return r;
455 }
456
457 if (!adev->gmc.xgmi.connected_to_cpu)
458 adev->gmc.xgmi.ras_funcs = &xgmi_ras_funcs;
459
460 if (adev->gmc.xgmi.ras_funcs &&
461 adev->gmc.xgmi.ras_funcs->ras_late_init) {
462 r = adev->gmc.xgmi.ras_funcs->ras_late_init(adev);
463 if (r)
464 return r;
465 }
466
467 if (adev->hdp.ras_funcs &&
468 adev->hdp.ras_funcs->ras_late_init) {
469 r = adev->hdp.ras_funcs->ras_late_init(adev);
470 if (r)
471 return r;
472 }
473
474 return 0;
475}
476
477void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
478{
479 if (adev->umc.ras_funcs &&
480 adev->umc.ras_funcs->ras_fini)
481 adev->umc.ras_funcs->ras_fini(adev);
482
483 if (adev->mmhub.ras_funcs &&
484 adev->mmhub.ras_funcs->ras_fini)
485 adev->mmhub.ras_funcs->ras_fini(adev);
486
487 if (adev->gmc.xgmi.ras_funcs &&
488 adev->gmc.xgmi.ras_funcs->ras_fini)
489 adev->gmc.xgmi.ras_funcs->ras_fini(adev);
490
491 if (adev->hdp.ras_funcs &&
492 adev->hdp.ras_funcs->ras_fini)
493 adev->hdp.ras_funcs->ras_fini(adev);
494}
495
496 /*
497 * The latest engine allocation on gfx9/10 is:
498 * Engine 2, 3: firmware
499 * Engine 0, 1, 4~16: amdgpu ring,
500 * subject to change when ring number changes
501 * Engine 17: Gart flushes
502 */
503#define GFXHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
504#define MMHUB_FREE_VM_INV_ENGS_BITMAP 0x1FFF3
505
506int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
507{
508 struct amdgpu_ring *ring;
509 unsigned vm_inv_engs[AMDGPU_MAX_VMHUBS] =
510 {GFXHUB_FREE_VM_INV_ENGS_BITMAP, MMHUB_FREE_VM_INV_ENGS_BITMAP,
511 GFXHUB_FREE_VM_INV_ENGS_BITMAP};
512 unsigned i;
513 unsigned vmhub, inv_eng;
514
515 for (i = 0; i < adev->num_rings; ++i) {
516 ring = adev->rings[i];
517 vmhub = ring->funcs->vmhub;
518
519 if (ring == &adev->mes.ring)
520 continue;
521
522 inv_eng = ffs(vm_inv_engs[vmhub]);
523 if (!inv_eng) {
524 dev_err(adev->dev, "no VM inv eng for ring %s\n",
525 ring->name);
526 return -EINVAL;
527 }
528
529 ring->vm_inv_eng = inv_eng - 1;
530 vm_inv_engs[vmhub] &= ~(1 << ring->vm_inv_eng);
531
532 dev_info(adev->dev, "ring %s uses VM inv eng %u on hub %u\n",
533 ring->name, ring->vm_inv_eng, ring->funcs->vmhub);
534 }
535
536 return 0;
537}
538
539/**
540 * amdgpu_gmc_tmz_set -- check and set if a device supports TMZ
541 * @adev: amdgpu_device pointer
542 *
543 * Check and set if an the device @adev supports Trusted Memory
544 * Zones (TMZ).
545 */
546void amdgpu_gmc_tmz_set(struct amdgpu_device *adev)
547{
548 switch (adev->asic_type) {
549 case CHIP_RAVEN:
550 case CHIP_RENOIR:
551 if (amdgpu_tmz == 0) {
552 adev->gmc.tmz_enabled = false;
553 dev_info(adev->dev,
554 "Trusted Memory Zone (TMZ) feature disabled (cmd line)\n");
555 } else {
556 adev->gmc.tmz_enabled = true;
557 dev_info(adev->dev,
558 "Trusted Memory Zone (TMZ) feature enabled\n");
559 }
560 break;
561 case CHIP_NAVI10:
562 case CHIP_NAVI14:
563 case CHIP_NAVI12:
564 case CHIP_VANGOGH:
565 case CHIP_YELLOW_CARP:
566 /* Don't enable it by default yet.
567 */
568 if (amdgpu_tmz < 1) {
569 adev->gmc.tmz_enabled = false;
570 dev_info(adev->dev,
571 "Trusted Memory Zone (TMZ) feature disabled as experimental (default)\n");
572 } else {
573 adev->gmc.tmz_enabled = true;
574 dev_info(adev->dev,
575 "Trusted Memory Zone (TMZ) feature enabled as experimental (cmd line)\n");
576 }
577 break;
578 default:
579 adev->gmc.tmz_enabled = false;
580 dev_warn(adev->dev,
581 "Trusted Memory Zone (TMZ) feature not supported\n");
582 break;
583 }
584}
585
586/**
587 * amdgpu_gmc_noretry_set -- set per asic noretry defaults
588 * @adev: amdgpu_device pointer
589 *
590 * Set a per asic default for the no-retry parameter.
591 *
592 */
593void amdgpu_gmc_noretry_set(struct amdgpu_device *adev)
594{
595 struct amdgpu_gmc *gmc = &adev->gmc;
596
597 switch (adev->asic_type) {
598 case CHIP_VEGA10:
599 case CHIP_VEGA20:
600 case CHIP_ARCTURUS:
601 case CHIP_ALDEBARAN:
602 /*
603 * noretry = 0 will cause kfd page fault tests fail
604 * for some ASICs, so set default to 1 for these ASICs.
605 */
606 if (amdgpu_noretry == -1)
607 gmc->noretry = 1;
608 else
609 gmc->noretry = amdgpu_noretry;
610 break;
611 case CHIP_RAVEN:
612 default:
613 /* Raven currently has issues with noretry
614 * regardless of what we decide for other
615 * asics, we should leave raven with
616 * noretry = 0 until we root cause the
617 * issues.
618 *
619 * default this to 0 for now, but we may want
620 * to change this in the future for certain
621 * GPUs as it can increase performance in
622 * certain cases.
623 */
624 if (amdgpu_noretry == -1)
625 gmc->noretry = 0;
626 else
627 gmc->noretry = amdgpu_noretry;
628 break;
629 }
630}
631
632void amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
633 bool enable)
634{
635 struct amdgpu_vmhub *hub;
636 u32 tmp, reg, i;
637
638 hub = &adev->vmhub[hub_type];
639 for (i = 0; i < 16; i++) {
640 reg = hub->vm_context0_cntl + hub->ctx_distance * i;
641
642 tmp = (hub_type == AMDGPU_GFXHUB_0) ?
643 RREG32_SOC15_IP(GC, reg) :
644 RREG32_SOC15_IP(MMHUB, reg);
645
646 if (enable)
647 tmp |= hub->vm_cntx_cntl_vm_fault;
648 else
649 tmp &= ~hub->vm_cntx_cntl_vm_fault;
650
651 (hub_type == AMDGPU_GFXHUB_0) ?
652 WREG32_SOC15_IP(GC, reg, tmp) :
653 WREG32_SOC15_IP(MMHUB, reg, tmp);
654 }
655}
656
657void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev)
658{
659 unsigned size;
660
661 /*
662 * TODO:
663 * Currently there is a bug where some memory client outside
664 * of the driver writes to first 8M of VRAM on S3 resume,
665 * this overrides GART which by default gets placed in first 8M and
666 * causes VM_FAULTS once GTT is accessed.
667 * Keep the stolen memory reservation until the while this is not solved.
668 */
669 switch (adev->asic_type) {
670 case CHIP_VEGA10:
671 case CHIP_RAVEN:
672 case CHIP_RENOIR:
673 adev->mman.keep_stolen_vga_memory = true;
674 break;
675 default:
676 adev->mman.keep_stolen_vga_memory = false;
677 break;
678 }
679
680 if (amdgpu_sriov_vf(adev) ||
681 !amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_DCE)) {
682 size = 0;
683 } else {
684 size = amdgpu_gmc_get_vbios_fb_size(adev);
685
686 if (adev->mman.keep_stolen_vga_memory)
687 size = max(size, (unsigned)AMDGPU_VBIOS_VGA_ALLOCATION);
688 }
689
690 /* set to 0 if the pre-OS buffer uses up most of vram */
691 if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
692 size = 0;
693
694 if (size > AMDGPU_VBIOS_VGA_ALLOCATION) {
695 adev->mman.stolen_vga_size = AMDGPU_VBIOS_VGA_ALLOCATION;
696 adev->mman.stolen_extended_size = size - adev->mman.stolen_vga_size;
697 } else {
698 adev->mman.stolen_vga_size = size;
699 adev->mman.stolen_extended_size = 0;
700 }
701}
702
703/**
704 * amdgpu_gmc_init_pdb0 - initialize PDB0
705 *
706 * @adev: amdgpu_device pointer
707 *
708 * This function is only used when GART page table is used
709 * for FB address translatioin. In such a case, we construct
710 * a 2-level system VM page table: PDB0->PTB, to cover both
711 * VRAM of the hive and system memory.
712 *
713 * PDB0 is static, initialized once on driver initialization.
714 * The first n entries of PDB0 are used as PTE by setting
715 * P bit to 1, pointing to VRAM. The n+1'th entry points
716 * to a big PTB covering system memory.
717 *
718 */
719void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev)
720{
721 int i;
722 uint64_t flags = adev->gart.gart_pte_flags; //TODO it is UC. explore NC/RW?
723 /* Each PDE0 (used as PTE) covers (2^vmid0_page_table_block_size)*2M
724 */
725 u64 vram_size = adev->gmc.xgmi.node_segment_size * adev->gmc.xgmi.num_physical_nodes;
726 u64 pde0_page_size = (1ULL<<adev->gmc.vmid0_page_table_block_size)<<21;
727 u64 vram_addr = adev->vm_manager.vram_base_offset -
728 adev->gmc.xgmi.physical_node_id * adev->gmc.xgmi.node_segment_size;
729 u64 vram_end = vram_addr + vram_size;
730 u64 gart_ptb_gpu_pa = amdgpu_gmc_vram_pa(adev, adev->gart.bo);
731
732 flags |= AMDGPU_PTE_VALID | AMDGPU_PTE_READABLE;
733 flags |= AMDGPU_PTE_WRITEABLE;
734 flags |= AMDGPU_PTE_SNOOPED;
735 flags |= AMDGPU_PTE_FRAG((adev->gmc.vmid0_page_table_block_size + 9*1));
736 flags |= AMDGPU_PDE_PTE;
737
738 /* The first n PDE0 entries are used as PTE,
739 * pointing to vram
740 */
741 for (i = 0; vram_addr < vram_end; i++, vram_addr += pde0_page_size)
742 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, vram_addr, flags);
743
744 /* The n+1'th PDE0 entry points to a huge
745 * PTB who has more than 512 entries each
746 * pointing to a 4K system page
747 */
748 flags = AMDGPU_PTE_VALID;
749 flags |= AMDGPU_PDE_BFS(0) | AMDGPU_PTE_SNOOPED;
750 /* Requires gart_ptb_gpu_pa to be 4K aligned */
751 amdgpu_gmc_set_pte_pde(adev, adev->gmc.ptr_pdb0, i, gart_ptb_gpu_pa, flags);
752}
753
754/**
755 * amdgpu_gmc_vram_mc2pa - calculate vram buffer's physical address from MC
756 * address
757 *
758 * @adev: amdgpu_device pointer
759 * @mc_addr: MC address of buffer
760 */
761uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr)
762{
763 return mc_addr - adev->gmc.vram_start + adev->vm_manager.vram_base_offset;
764}
765
766/**
767 * amdgpu_gmc_vram_pa - calculate vram buffer object's physical address from
768 * GPU's view
769 *
770 * @adev: amdgpu_device pointer
771 * @bo: amdgpu buffer object
772 */
773uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
774{
775 return amdgpu_gmc_vram_mc2pa(adev, amdgpu_bo_gpu_offset(bo));
776}
777
778/**
779 * amdgpu_gmc_vram_cpu_pa - calculate vram buffer object's physical address
780 * from CPU's view
781 *
782 * @adev: amdgpu_device pointer
783 * @bo: amdgpu buffer object
784 */
785uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo)
786{
787 return amdgpu_bo_gpu_offset(bo) - adev->gmc.vram_start + adev->gmc.aper_base;
788}
789
790void amdgpu_gmc_get_reserved_allocation(struct amdgpu_device *adev)
791{
792 /* Some ASICs need to reserve a region of video memory to avoid access
793 * from driver */
794 adev->mman.stolen_reserved_offset = 0;
795 adev->mman.stolen_reserved_size = 0;
796
797 switch (adev->asic_type) {
798 case CHIP_YELLOW_CARP:
799 if (amdgpu_discovery == 0) {
800 adev->mman.stolen_reserved_offset = 0x1ffb0000;
801 adev->mman.stolen_reserved_size = 64 * PAGE_SIZE;
802 }
803 break;
804 default:
805 break;
806 }
807}