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 /*
67  * GMC page fault information
68  */
69 struct amdgpu_gmc_fault {
70 	uint64_t	timestamp:48;
71 	uint64_t	next:AMDGPU_GMC_FAULT_RING_ORDER;
72 	atomic64_t	key;
73 	uint64_t	timestamp_expiry:48;
74 };
75 
76 /*
77  * VMHUB structures, functions & helpers
78  */
79 struct amdgpu_vmhub_funcs {
80 	void (*print_l2_protection_fault_status)(struct amdgpu_device *adev,
81 						 uint32_t status);
82 	uint32_t (*get_invalidate_req)(unsigned int vmid, uint32_t flush_type);
83 };
84 
85 struct amdgpu_vmhub {
86 	uint32_t	ctx0_ptb_addr_lo32;
87 	uint32_t	ctx0_ptb_addr_hi32;
88 	uint32_t	vm_inv_eng0_sem;
89 	uint32_t	vm_inv_eng0_req;
90 	uint32_t	vm_inv_eng0_ack;
91 	uint32_t	vm_context0_cntl;
92 	uint32_t	vm_l2_pro_fault_status;
93 	uint32_t	vm_l2_pro_fault_cntl;
94 
95 	/*
96 	 * store the register distances between two continuous context domain
97 	 * and invalidation engine.
98 	 */
99 	uint32_t	ctx_distance;
100 	uint32_t	ctx_addr_distance; /* include LO32/HI32 */
101 	uint32_t	eng_distance;
102 	uint32_t	eng_addr_distance; /* include LO32/HI32 */
103 
104 	uint32_t        vm_cntx_cntl;
105 	uint32_t	vm_cntx_cntl_vm_fault;
106 	uint32_t	vm_l2_bank_select_reserved_cid2;
107 
108 	uint32_t	vm_contexts_disable;
109 
110 	const struct amdgpu_vmhub_funcs *vmhub_funcs;
111 };
112 
113 /*
114  * GPU MC structures, functions & helpers
115  */
116 struct amdgpu_gmc_funcs {
117 	/* flush the vm tlb via mmio */
118 	void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid,
119 				uint32_t vmhub, uint32_t flush_type);
120 	/* flush the vm tlb via pasid */
121 	int (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid,
122 					uint32_t flush_type, bool all_hub);
123 	/* flush the vm tlb via ring */
124 	uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid,
125 				       uint64_t pd_addr);
126 	/* Change the VMID -> PASID mapping */
127 	void (*emit_pasid_mapping)(struct amdgpu_ring *ring, unsigned vmid,
128 				   unsigned pasid);
129 	/* enable/disable PRT support */
130 	void (*set_prt)(struct amdgpu_device *adev, bool enable);
131 	/* map mtype to hardware flags */
132 	uint64_t (*map_mtype)(struct amdgpu_device *adev, uint32_t flags);
133 	/* get the pde for a given mc addr */
134 	void (*get_vm_pde)(struct amdgpu_device *adev, int level,
135 			   u64 *dst, u64 *flags);
136 	/* get the pte flags to use for a BO VA mapping */
137 	void (*get_vm_pte)(struct amdgpu_device *adev,
138 			   struct amdgpu_bo_va_mapping *mapping,
139 			   uint64_t *flags);
140 	/* get the amount of memory used by the vbios for pre-OS console */
141 	unsigned int (*get_vbios_fb_size)(struct amdgpu_device *adev);
142 };
143 
144 struct amdgpu_xgmi_ras {
145 	struct amdgpu_ras_block_object ras_block;
146 };
147 
148 struct amdgpu_xgmi {
149 	/* from psp */
150 	u64 node_id;
151 	u64 hive_id;
152 	/* fixed per family */
153 	u64 node_segment_size;
154 	/* physical node (0-3) */
155 	unsigned physical_node_id;
156 	/* number of nodes (0-4) */
157 	unsigned num_physical_nodes;
158 	/* gpu list in the same hive */
159 	struct list_head head;
160 	bool supported;
161 	struct ras_common_if *ras_if;
162 	bool connected_to_cpu;
163 	bool pending_reset;
164 	struct amdgpu_xgmi_ras *ras;
165 };
166 
167 struct amdgpu_gmc {
168 	/* FB's physical address in MMIO space (for CPU to
169 	 * map FB). This is different compared to the agp/
170 	 * gart/vram_start/end field as the later is from
171 	 * GPU's view and aper_base is from CPU's view.
172 	 */
173 	resource_size_t		aper_size;
174 	resource_size_t		aper_base;
175 	/* for some chips with <= 32MB we need to lie
176 	 * about vram size near mc fb location */
177 	u64			mc_vram_size;
178 	u64			visible_vram_size;
179 	/* AGP aperture start and end in MC address space
180 	 * Driver find a hole in the MC address space
181 	 * to place AGP by setting MC_VM_AGP_BOT/TOP registers
182 	 * Under VMID0, logical address == MC address. AGP
183 	 * aperture maps to physical bus or IOVA addressed.
184 	 * AGP aperture is used to simulate FB in ZFB case.
185 	 * AGP aperture is also used for page table in system
186 	 * memory (mainly for APU).
187 	 *
188 	 */
189 	u64			agp_size;
190 	u64			agp_start;
191 	u64			agp_end;
192 	/* GART aperture start and end in MC address space
193 	 * Driver find a hole in the MC address space
194 	 * to place GART by setting VM_CONTEXT0_PAGE_TABLE_START/END_ADDR
195 	 * registers
196 	 * Under VMID0, logical address inside GART aperture will
197 	 * be translated through gpuvm gart page table to access
198 	 * paged system memory
199 	 */
200 	u64			gart_size;
201 	u64			gart_start;
202 	u64			gart_end;
203 	/* Frame buffer aperture of this GPU device. Different from
204 	 * fb_start (see below), this only covers the local GPU device.
205 	 * If driver uses FB aperture to access FB, driver get fb_start from
206 	 * MC_VM_FB_LOCATION_BASE (set by vbios) and calculate vram_start
207 	 * of this local device by adding an offset inside the XGMI hive.
208 	 * If driver uses GART table for VMID0 FB access, driver finds a hole in
209 	 * VMID0's virtual address space to place the SYSVM aperture inside
210 	 * which the first part is vram and the second part is gart (covering
211 	 * system ram).
212 	 */
213 	u64			vram_start;
214 	u64			vram_end;
215 	/* FB region , it's same as local vram region in single GPU, in XGMI
216 	 * configuration, this region covers all GPUs in the same hive ,
217 	 * each GPU in the hive has the same view of this FB region .
218 	 * GPU0's vram starts at offset (0 * segment size) ,
219 	 * GPU1 starts at offset (1 * segment size), etc.
220 	 */
221 	u64			fb_start;
222 	u64			fb_end;
223 	unsigned		vram_width;
224 	u64			real_vram_size;
225 	int			vram_mtrr;
226 	u64                     mc_mask;
227 	const struct firmware   *fw;	/* MC firmware */
228 	uint32_t                fw_version;
229 	struct amdgpu_irq_src	vm_fault;
230 	uint32_t		vram_type;
231 	uint8_t			vram_vendor;
232 	uint32_t                srbm_soft_reset;
233 	bool			prt_warning;
234 	uint32_t		sdpif_register;
235 	/* apertures */
236 	u64			shared_aperture_start;
237 	u64			shared_aperture_end;
238 	u64			private_aperture_start;
239 	u64			private_aperture_end;
240 	/* protects concurrent invalidation */
241 	spinlock_t		invalidate_lock;
242 	bool			translate_further;
243 	struct kfd_vm_fault_info *vm_fault_info;
244 	atomic_t		vm_fault_info_updated;
245 
246 	struct amdgpu_gmc_fault	fault_ring[AMDGPU_GMC_FAULT_RING_SIZE];
247 	struct {
248 		uint64_t	idx:AMDGPU_GMC_FAULT_RING_ORDER;
249 	} fault_hash[AMDGPU_GMC_FAULT_HASH_SIZE];
250 	uint64_t		last_fault:AMDGPU_GMC_FAULT_RING_ORDER;
251 
252 	bool tmz_enabled;
253 
254 	const struct amdgpu_gmc_funcs	*gmc_funcs;
255 
256 	struct amdgpu_xgmi xgmi;
257 	struct amdgpu_irq_src	ecc_irq;
258 	int noretry;
259 
260 	uint32_t	vmid0_page_table_block_size;
261 	uint32_t	vmid0_page_table_depth;
262 	struct amdgpu_bo		*pdb0_bo;
263 	/* CPU kmapped address of pdb0*/
264 	void				*ptr_pdb0;
265 
266 	/* MALL size */
267 	u64 mall_size;
268 	/* number of UMC instances */
269 	int num_umc;
270 	/* mode2 save restore */
271 	u64 VM_L2_CNTL;
272 	u64 VM_L2_CNTL2;
273 	u64 VM_DUMMY_PAGE_FAULT_CNTL;
274 	u64 VM_DUMMY_PAGE_FAULT_ADDR_LO32;
275 	u64 VM_DUMMY_PAGE_FAULT_ADDR_HI32;
276 	u64 VM_L2_PROTECTION_FAULT_CNTL;
277 	u64 VM_L2_PROTECTION_FAULT_CNTL2;
278 	u64 VM_L2_PROTECTION_FAULT_MM_CNTL3;
279 	u64 VM_L2_PROTECTION_FAULT_MM_CNTL4;
280 	u64 VM_L2_PROTECTION_FAULT_ADDR_LO32;
281 	u64 VM_L2_PROTECTION_FAULT_ADDR_HI32;
282 	u64 VM_DEBUG;
283 	u64 VM_L2_MM_GROUP_RT_CLASSES;
284 	u64 VM_L2_BANK_SELECT_RESERVED_CID;
285 	u64 VM_L2_BANK_SELECT_RESERVED_CID2;
286 	u64 VM_L2_CACHE_PARITY_CNTL;
287 	u64 VM_L2_IH_LOG_CNTL;
288 	u64 VM_CONTEXT_CNTL[16];
289 	u64 VM_CONTEXT_PAGE_TABLE_BASE_ADDR_LO32[16];
290 	u64 VM_CONTEXT_PAGE_TABLE_BASE_ADDR_HI32[16];
291 	u64 VM_CONTEXT_PAGE_TABLE_START_ADDR_LO32[16];
292 	u64 VM_CONTEXT_PAGE_TABLE_START_ADDR_HI32[16];
293 	u64 VM_CONTEXT_PAGE_TABLE_END_ADDR_LO32[16];
294 	u64 VM_CONTEXT_PAGE_TABLE_END_ADDR_HI32[16];
295 	u64 MC_VM_MX_L1_TLB_CNTL;
296 };
297 
298 #define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type)))
299 #define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub) \
300 	((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \
301 	((adev), (pasid), (type), (allhub)))
302 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr))
303 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid))
304 #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags))
305 #define amdgpu_gmc_get_vm_pde(adev, level, dst, flags) (adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (dst), (flags))
306 #define amdgpu_gmc_get_vm_pte(adev, mapping, flags) (adev)->gmc.gmc_funcs->get_vm_pte((adev), (mapping), (flags))
307 #define amdgpu_gmc_get_vbios_fb_size(adev) (adev)->gmc.gmc_funcs->get_vbios_fb_size((adev))
308 
309 /**
310  * amdgpu_gmc_vram_full_visible - Check if full VRAM is visible through the BAR
311  *
312  * @adev: amdgpu_device pointer
313  *
314  * Returns:
315  * True if full VRAM is visible through the BAR
316  */
317 static inline bool amdgpu_gmc_vram_full_visible(struct amdgpu_gmc *gmc)
318 {
319 	WARN_ON(gmc->real_vram_size < gmc->visible_vram_size);
320 
321 	return (gmc->real_vram_size == gmc->visible_vram_size);
322 }
323 
324 /**
325  * amdgpu_gmc_sign_extend - sign extend the given gmc address
326  *
327  * @addr: address to extend
328  */
329 static inline uint64_t amdgpu_gmc_sign_extend(uint64_t addr)
330 {
331 	if (addr >= AMDGPU_GMC_HOLE_START)
332 		addr |= AMDGPU_GMC_HOLE_END;
333 
334 	return addr;
335 }
336 
337 int amdgpu_gmc_pdb0_alloc(struct amdgpu_device *adev);
338 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
339 			       uint64_t *addr, uint64_t *flags);
340 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
341 				uint32_t gpu_page_idx, uint64_t addr,
342 				uint64_t flags);
343 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo);
344 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo);
345 void amdgpu_gmc_sysvm_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc);
346 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
347 			      u64 base);
348 void amdgpu_gmc_gart_location(struct amdgpu_device *adev,
349 			      struct amdgpu_gmc *mc);
350 void amdgpu_gmc_agp_location(struct amdgpu_device *adev,
351 			     struct amdgpu_gmc *mc);
352 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev,
353 			      struct amdgpu_ih_ring *ih, uint64_t addr,
354 			      uint16_t pasid, uint64_t timestamp);
355 void amdgpu_gmc_filter_faults_remove(struct amdgpu_device *adev, uint64_t addr,
356 				     uint16_t pasid);
357 int amdgpu_gmc_ras_sw_init(struct amdgpu_device *adev);
358 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
359 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
360 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
361 
362 extern void amdgpu_gmc_tmz_set(struct amdgpu_device *adev);
363 extern void amdgpu_gmc_noretry_set(struct amdgpu_device *adev);
364 
365 extern void
366 amdgpu_gmc_set_vm_fault_masks(struct amdgpu_device *adev, int hub_type,
367 			      bool enable);
368 
369 void amdgpu_gmc_get_vbios_allocations(struct amdgpu_device *adev);
370 
371 void amdgpu_gmc_init_pdb0(struct amdgpu_device *adev);
372 uint64_t amdgpu_gmc_vram_mc2pa(struct amdgpu_device *adev, uint64_t mc_addr);
373 uint64_t amdgpu_gmc_vram_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo);
374 uint64_t amdgpu_gmc_vram_cpu_pa(struct amdgpu_device *adev, struct amdgpu_bo *bo);
375 int amdgpu_gmc_vram_checking(struct amdgpu_device *adev);
376 #endif
377