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 #ifndef __AMDGPU_GMC_H__
27 #define __AMDGPU_GMC_H__
28
29 #include <linux/types.h>
30
31 #include "amdgpu_irq.h"
32 #include "amdgpu_ras.h"
33
34 /* VA hole for 48bit addresses on Vega10 */
35 #define AMDGPU_GMC_HOLE_START 0x0000800000000000ULL
36 #define AMDGPU_GMC_HOLE_END 0xffff800000000000ULL
37
38 /*
39 * Hardware is programmed as if the hole doesn't exists with start and end
40 * address values.
41 *
42 * This mask is used to remove the upper 16bits of the VA and so come up with
43 * the linear addr value.
44 */
45 #define AMDGPU_GMC_HOLE_MASK 0x0000ffffffffffffULL
46
47 /*
48 * Ring size as power of two for the log of recent faults.
49 */
50 #define AMDGPU_GMC_FAULT_RING_ORDER 8
51 #define AMDGPU_GMC_FAULT_RING_SIZE (1 << AMDGPU_GMC_FAULT_RING_ORDER)
52
53 /*
54 * Hash size as power of two for the log of recent faults
55 */
56 #define AMDGPU_GMC_FAULT_HASH_ORDER 8
57 #define AMDGPU_GMC_FAULT_HASH_SIZE (1 << AMDGPU_GMC_FAULT_HASH_ORDER)
58
59 /*
60 * Number of IH timestamp ticks until a fault is considered handled
61 */
62 #define AMDGPU_GMC_FAULT_TIMEOUT 5000ULL
63
64 struct firmware;
65
66 enum amdgpu_memory_partition {
67 UNKNOWN_MEMORY_PARTITION_MODE = 0,
68 AMDGPU_NPS1_PARTITION_MODE = 1,
69 AMDGPU_NPS2_PARTITION_MODE = 2,
70 AMDGPU_NPS3_PARTITION_MODE = 3,
71 AMDGPU_NPS4_PARTITION_MODE = 4,
72 AMDGPU_NPS6_PARTITION_MODE = 6,
73 AMDGPU_NPS8_PARTITION_MODE = 8,
74 };
75
76 #define AMDGPU_ALL_NPS_MASK \
77 (BIT(AMDGPU_NPS1_PARTITION_MODE) | BIT(AMDGPU_NPS2_PARTITION_MODE) | \
78 BIT(AMDGPU_NPS3_PARTITION_MODE) | BIT(AMDGPU_NPS4_PARTITION_MODE) | \
79 BIT(AMDGPU_NPS6_PARTITION_MODE) | BIT(AMDGPU_NPS8_PARTITION_MODE))
80
81 #define AMDGPU_GMC_INIT_RESET_NPS BIT(0)
82
83 /*
84 * GMC page fault information
85 */
86 struct amdgpu_gmc_fault {
87 uint64_t timestamp:48;
88 uint64_t next:AMDGPU_GMC_FAULT_RING_ORDER;
89 atomic64_t key;
90 uint64_t timestamp_expiry:48;
91 };
92
93 /*
94 * VMHUB structures, functions & helpers
95 */
96 struct amdgpu_vmhub_funcs {
97 void (*print_l2_protection_fault_status)(struct amdgpu_device *adev,
98 uint32_t status);
99 uint32_t (*get_invalidate_req)(unsigned int vmid, uint32_t flush_type);
100 };
101
102 struct amdgpu_vmhub {
103 uint32_t ctx0_ptb_addr_lo32;
104 uint32_t ctx0_ptb_addr_hi32;
105 uint32_t vm_inv_eng0_sem;
106 uint32_t vm_inv_eng0_req;
107 uint32_t vm_inv_eng0_ack;
108 uint32_t vm_context0_cntl;
109 uint32_t vm_l2_pro_fault_status;
110 uint32_t vm_l2_pro_fault_cntl;
111
112 /*
113 * store the register distances between two continuous context domain
114 * and invalidation engine.
115 */
116 uint32_t ctx_distance;
117 uint32_t ctx_addr_distance; /* include LO32/HI32 */
118 uint32_t eng_distance;
119 uint32_t eng_addr_distance; /* include LO32/HI32 */
120
121 uint32_t vm_cntx_cntl;
122 uint32_t vm_cntx_cntl_vm_fault;
123 uint32_t vm_l2_bank_select_reserved_cid2;
124
125 uint32_t vm_contexts_disable;
126
127 bool sdma_invalidation_workaround;
128
129 const struct amdgpu_vmhub_funcs *vmhub_funcs;
130 };
131
132 /*
133 * GPU MC structures, functions & helpers
134 */
135 struct amdgpu_gmc_funcs {
136 /* flush the vm tlb via mmio */
137 void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid,
138 uint32_t vmhub, uint32_t flush_type);
139 /* flush the vm tlb via pasid */
140 void (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid,
141 uint32_t flush_type, bool all_hub,
142 uint32_t inst);
143 /* flush the vm tlb via ring */
144 uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid,
145 uint64_t pd_addr);
146 /* Change the VMID -> PASID mapping */
147 void (*emit_pasid_mapping)(struct amdgpu_ring *ring, unsigned vmid,
148 unsigned pasid);
149 /* enable/disable PRT support */
150 void (*set_prt)(struct amdgpu_device *adev, bool enable);
151 /* map mtype to hardware flags */
152 uint64_t (*map_mtype)(struct amdgpu_device *adev, uint32_t flags);
153 /* get the pde for a given mc addr */
154 void (*get_vm_pde)(struct amdgpu_device *adev, int level,
155 u64 *dst, u64 *flags);
156 /* get the pte flags to use for a BO VA mapping */
157 void (*get_vm_pte)(struct amdgpu_device *adev,
158 struct amdgpu_bo_va_mapping *mapping,
159 uint64_t *flags);
160 /* override per-page pte flags */
161 void (*override_vm_pte_flags)(struct amdgpu_device *dev,
162 struct amdgpu_vm *vm,
163 uint64_t addr, uint64_t *flags);
164 /* get the amount of memory used by the vbios for pre-OS console */
165 unsigned int (*get_vbios_fb_size)(struct amdgpu_device *adev);
166 /* get the DCC buffer alignment */
167 unsigned int (*get_dcc_alignment)(struct amdgpu_device *adev);
168
169 enum amdgpu_memory_partition (*query_mem_partition_mode)(
170 struct amdgpu_device *adev);
171 /* Request NPS mode */
172 int (*request_mem_partition_mode)(struct amdgpu_device *adev,
173 int nps_mode);
174 bool (*need_reset_on_init)(struct amdgpu_device *adev);
175 };
176
177 struct amdgpu_xgmi_ras {
178 struct amdgpu_ras_block_object ras_block;
179 };
180
181 struct amdgpu_xgmi {
182 /* from psp */
183 u64 node_id;
184 u64 hive_id;
185 /* fixed per family */
186 u64 node_segment_size;
187 /* physical node (0-3) */
188 unsigned physical_node_id;
189 /* number of nodes (0-4) */
190 unsigned num_physical_nodes;
191 /* gpu list in the same hive */
192 struct list_head head;
193 bool supported;
194 struct ras_common_if *ras_if;
195 bool connected_to_cpu;
196 struct amdgpu_xgmi_ras *ras;
197 };
198
199 struct amdgpu_mem_partition_info {
200 union {
201 struct {
202 uint32_t fpfn;
203 uint32_t lpfn;
204 } range;
205 struct {
206 int node;
207 } numa;
208 };
209 uint64_t size;
210 };
211
212 #define INVALID_PFN -1
213
214 struct amdgpu_gmc_memrange {
215 uint64_t base_address;
216 uint64_t limit_address;
217 uint32_t flags;
218 int nid_mask;
219 };
220
221 enum amdgpu_gart_placement {
222 AMDGPU_GART_PLACEMENT_BEST_FIT = 0,
223 AMDGPU_GART_PLACEMENT_HIGH,
224 AMDGPU_GART_PLACEMENT_LOW,
225 };
226
227 struct amdgpu_gmc {
228 /* FB's physical address in MMIO space (for CPU to
229 * map FB). This is different compared to the agp/
230 * gart/vram_start/end field as the later is from
231 * GPU's view and aper_base is from CPU's view.
232 */
233 resource_size_t aper_size;
234 resource_size_t aper_base;
235 /* for some chips with <= 32MB we need to lie
236 * about vram size near mc fb location */
237 u64 mc_vram_size;
238 u64 visible_vram_size;
239 /* AGP aperture start and end in MC address space
240 * Driver find a hole in the MC address space
241 * to place AGP by setting MC_VM_AGP_BOT/TOP registers
242 * Under VMID0, logical address == MC address. AGP
243 * aperture maps to physical bus or IOVA addressed.
244 * AGP aperture is used to simulate FB in ZFB case.
245 * AGP aperture is also used for page table in system
246 * memory (mainly for APU).
247 *
248 */
249 u64 agp_size;
250 u64 agp_start;
251 u64 agp_end;
252 /* GART aperture start and end in MC address space
253 * Driver find a hole in the MC address space
254 * to place GART by setting VM_CONTEXT0_PAGE_TABLE_START/END_ADDR
255 * registers
256 * Under VMID0, logical address inside GART aperture will
257 * be translated through gpuvm gart page table to access
258 * paged system memory
259 */
260 u64 gart_size;
261 u64 gart_start;
262 u64 gart_end;
263 /* Frame buffer aperture of this GPU device. Different from
264 * fb_start (see below), this only covers the local GPU device.
265 * If driver uses FB aperture to access FB, driver get fb_start from
266 * MC_VM_FB_LOCATION_BASE (set by vbios) and calculate vram_start
267 * of this local device by adding an offset inside the XGMI hive.
268 * If driver uses GART table for VMID0 FB access, driver finds a hole in
269 * VMID0's virtual address space to place the SYSVM aperture inside
270 * which the first part is vram and the second part is gart (covering
271 * system ram).
272 */
273 u64 vram_start;
274 u64 vram_end;
275 /* FB region , it's same as local vram region in single GPU, in XGMI
276 * configuration, this region covers all GPUs in the same hive ,
277 * each GPU in the hive has the same view of this FB region .
278 * GPU0's vram starts at offset (0 * segment size) ,
279 * GPU1 starts at offset (1 * segment size), etc.
280 */
281 u64 fb_start;
282 u64 fb_end;
283 unsigned vram_width;
284 u64 real_vram_size;
285 int vram_mtrr;
286 u64 mc_mask;
287 const struct firmware *fw; /* MC firmware */
288 uint32_t fw_version;
289 struct amdgpu_irq_src vm_fault;
290 uint32_t vram_type;
291 uint8_t vram_vendor;
292 uint32_t srbm_soft_reset;
293 bool prt_warning;
294 uint32_t sdpif_register;
295 /* apertures */
296 u64 shared_aperture_start;
297 u64 shared_aperture_end;
298 u64 private_aperture_start;
299 u64 private_aperture_end;
300 /* protects concurrent invalidation */
301 spinlock_t invalidate_lock;
302 bool translate_further;
303 struct kfd_vm_fault_info *vm_fault_info;
304 atomic_t vm_fault_info_updated;
305
306 struct amdgpu_gmc_fault fault_ring[AMDGPU_GMC_FAULT_RING_SIZE];
307 struct {
308 uint64_t idx:AMDGPU_GMC_FAULT_RING_ORDER;
309 } fault_hash[AMDGPU_GMC_FAULT_HASH_SIZE];
310 uint64_t last_fault:AMDGPU_GMC_FAULT_RING_ORDER;
311
312 bool tmz_enabled;
313 bool is_app_apu;
314
315 struct amdgpu_mem_partition_info *mem_partitions;
316 uint8_t num_mem_partitions;
317 const struct amdgpu_gmc_funcs *gmc_funcs;
318 enum amdgpu_memory_partition requested_nps_mode;
319 uint32_t supported_nps_modes;
320 uint32_t reset_flags;
321
322 struct amdgpu_xgmi xgmi;
323 struct amdgpu_irq_src ecc_irq;
324 int noretry;
325
326 uint32_t vmid0_page_table_block_size;
327 uint32_t vmid0_page_table_depth;
328 struct amdgpu_bo *pdb0_bo;
329 /* CPU kmapped address of pdb0*/
330 void *ptr_pdb0;
331
332 /* MALL size */
333 u64 mall_size;
334 uint32_t m_half_use;
335
336 /* number of UMC instances */
337 int num_umc;
338 /* mode2 save restore */
339 u64 VM_L2_CNTL;
340 u64 VM_L2_CNTL2;
341 u64 VM_DUMMY_PAGE_FAULT_CNTL;
342 u64 VM_DUMMY_PAGE_FAULT_ADDR_LO32;
343 u64 VM_DUMMY_PAGE_FAULT_ADDR_HI32;
344 u64 VM_L2_PROTECTION_FAULT_CNTL;
345 u64 VM_L2_PROTECTION_FAULT_CNTL2;
346 u64 VM_L2_PROTECTION_FAULT_MM_CNTL3;
347 u64 VM_L2_PROTECTION_FAULT_MM_CNTL4;
348 u64 VM_L2_PROTECTION_FAULT_ADDR_LO32;
349 u64 VM_L2_PROTECTION_FAULT_ADDR_HI32;
350 u64 VM_DEBUG;
351 u64 VM_L2_MM_GROUP_RT_CLASSES;
352 u64 VM_L2_BANK_SELECT_RESERVED_CID;
353 u64 VM_L2_BANK_SELECT_RESERVED_CID2;
354 u64 VM_L2_CACHE_PARITY_CNTL;
355 u64 VM_L2_IH_LOG_CNTL;
356 u64 VM_CONTEXT_CNTL[16];
357 u64 VM_CONTEXT_PAGE_TABLE_BASE_ADDR_LO32[16];
358 u64 VM_CONTEXT_PAGE_TABLE_BASE_ADDR_HI32[16];
359 u64 VM_CONTEXT_PAGE_TABLE_START_ADDR_LO32[16];
360 u64 VM_CONTEXT_PAGE_TABLE_START_ADDR_HI32[16];
361 u64 VM_CONTEXT_PAGE_TABLE_END_ADDR_LO32[16];
362 u64 VM_CONTEXT_PAGE_TABLE_END_ADDR_HI32[16];
363 u64 MC_VM_MX_L1_TLB_CNTL;
364
365 u64 noretry_flags;
366
367 bool flush_tlb_needs_extra_type_0;
368 bool flush_tlb_needs_extra_type_2;
369 bool flush_pasid_uses_kiq;
370 };
371
372 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr))
373 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid))
374 #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags))
375 #define amdgpu_gmc_get_vm_pde(adev, level, dst, flags) (adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (dst), (flags))
376 #define amdgpu_gmc_get_vm_pte(adev, mapping, flags) (adev)->gmc.gmc_funcs->get_vm_pte((adev), (mapping), (flags))
377 #define amdgpu_gmc_override_vm_pte_flags(adev, vm, addr, pte_flags) \
378 (adev)->gmc.gmc_funcs->override_vm_pte_flags \
379 ((adev), (vm), (addr), (pte_flags))
380 #define amdgpu_gmc_get_vbios_fb_size(adev) (adev)->gmc.gmc_funcs->get_vbios_fb_size((adev))
381 #define amdgpu_gmc_get_dcc_alignment(adev) ({ \
382 typeof(adev) _adev = (adev); \
383 _adev->gmc.gmc_funcs->get_dcc_alignment(_adev); \
384 })
385
386 /**
387 * amdgpu_gmc_vram_full_visible - Check if full VRAM is visible through the BAR
388 *
389 * @adev: amdgpu_device pointer
390 *
391 * Returns:
392 * True if full VRAM is visible through the BAR
393 */
amdgpu_gmc_vram_full_visible(struct amdgpu_gmc * gmc)394 static inline bool amdgpu_gmc_vram_full_visible(struct amdgpu_gmc *gmc)
395 {
396 WARN_ON(gmc->real_vram_size < gmc->visible_vram_size);
397
398 return (gmc->real_vram_size == gmc->visible_vram_size);
399 }
400
401 /**
402 * amdgpu_gmc_sign_extend - sign extend the given gmc address
403 *
404 * @addr: address to extend
405 */
amdgpu_gmc_sign_extend(uint64_t addr)406 static inline uint64_t amdgpu_gmc_sign_extend(uint64_t addr)
407 {
408 if (addr >= AMDGPU_GMC_HOLE_START)
409 addr |= AMDGPU_GMC_HOLE_END;
410
411 return addr;
412 }
413
414 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev);
415 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
416 uint64_t *addr, uint64_t *flags);
417 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
418 uint32_t gpu_page_idx, uint64_t addr,
419 uint64_t flags);
420 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo);
421 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo);
422 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc);
423 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
424 u64 base);
425 void amdgpu_gmc_gart_location(struct amdgpu_device *adev,
426 struct amdgpu_gmc *mc,
427 enum amdgpu_gart_placement gart_placement);
428 void amdgpu_gmc_agp_location(struct amdgpu_device *adev,
429 struct amdgpu_gmc *mc);
430 void amdgpu_gmc_set_agp_default(struct amdgpu_device *adev,
431 struct amdgpu_gmc *mc);
432 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
433 struct amdgpu_ih_ring *ih, uint64_t addr,
434 uint16_t pasid, uint64_t timestamp);
435 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
436 uint16_t pasid);
437 int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev);
438 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
439 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
440 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
441 void amdgpu_gmc_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
442 uint32_t vmhub, uint32_t flush_type);
443 int amdgpu_gmc_flush_gpu_tlb_pasid(struct amdgpu_device *adev, uint16_t pasid,
444 uint32_t flush_type, bool all_hub,
445 uint32_t inst);
446 void amdgpu_gmc_fw_reg_write_reg_wait(struct amdgpu_device *adev,
447 uint32_t reg0, uint32_t reg1,
448 uint32_t ref, uint32_t mask,
449 uint32_t xcc_inst);
450
451 extern void amdgpu_gmc_tmz_set(struct amdgpu_device *adev);
452 extern void amdgpu_gmc_noretry_set(struct amdgpu_device *adev);
453
454 extern void
455 amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
456 bool enable);
457
458 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev);
459
460 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev);
461 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr);
462 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo);
463 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev);
464 int amdgpu_gmc_sysfs_init(struct amdgpu_device *adev);
465 void amdgpu_gmc_sysfs_fini(struct amdgpu_device *adev);
466
467 int amdgpu_gmc_get_nps_memranges(struct amdgpu_device *adev,
468 struct amdgpu_mem_partition_info *mem_ranges,
469 uint8_t *exp_ranges);
470
471 int amdgpu_gmc_request_memory_partition(struct amdgpu_device *adev,
472 int nps_mode);
473 void amdgpu_gmc_prepare_nps_mode_change(struct amdgpu_device *adev);
474 bool amdgpu_gmc_need_reset_on_init(struct amdgpu_device *adev);
475
476 #endif
477