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 
33 /* VA hole for 48bit addresses on Vega10 */
34 #define AMDGPU_GMC_HOLE_START	0x0000800000000000ULL
35 #define AMDGPU_GMC_HOLE_END	0xffff800000000000ULL
36 
37 /*
38  * Hardware is programmed as if the hole doesn't exists with start and end
39  * address values.
40  *
41  * This mask is used to remove the upper 16bits of the VA and so come up with
42  * the linear addr value.
43  */
44 #define AMDGPU_GMC_HOLE_MASK	0x0000ffffffffffffULL
45 
46 /*
47  * Ring size as power of two for the log of recent faults.
48  */
49 #define AMDGPU_GMC_FAULT_RING_ORDER	8
50 #define AMDGPU_GMC_FAULT_RING_SIZE	(1 << AMDGPU_GMC_FAULT_RING_ORDER)
51 
52 /*
53  * Hash size as power of two for the log of recent faults
54  */
55 #define AMDGPU_GMC_FAULT_HASH_ORDER	8
56 #define AMDGPU_GMC_FAULT_HASH_SIZE	(1 << AMDGPU_GMC_FAULT_HASH_ORDER)
57 
58 /*
59  * Number of IH timestamp ticks until a fault is considered handled
60  */
61 #define AMDGPU_GMC_FAULT_TIMEOUT	5000ULL
62 
63 struct firmware;
64 
65 /*
66  * GMC page fault information
67  */
68 struct amdgpu_gmc_fault {
69 	uint64_t	timestamp;
70 	uint64_t	next:AMDGPU_GMC_FAULT_RING_ORDER;
71 	uint64_t	key:52;
72 };
73 
74 /*
75  * VMHUB structures, functions & helpers
76  */
77 struct amdgpu_vmhub {
78 	uint32_t	ctx0_ptb_addr_lo32;
79 	uint32_t	ctx0_ptb_addr_hi32;
80 	uint32_t	vm_inv_eng0_sem;
81 	uint32_t	vm_inv_eng0_req;
82 	uint32_t	vm_inv_eng0_ack;
83 	uint32_t	vm_context0_cntl;
84 	uint32_t	vm_l2_pro_fault_status;
85 	uint32_t	vm_l2_pro_fault_cntl;
86 
87 	/*
88 	 * store the register distances between two continuous context domain
89 	 * and invalidation engine.
90 	 */
91 	uint32_t	ctx_distance;
92 	uint32_t	ctx_addr_distance; /* include LO32/HI32 */
93 	uint32_t	eng_distance;
94 	uint32_t	eng_addr_distance; /* include LO32/HI32 */
95 };
96 
97 /*
98  * GPU MC structures, functions & helpers
99  */
100 struct amdgpu_gmc_funcs {
101 	/* flush the vm tlb via mmio */
102 	void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid,
103 				uint32_t vmhub, uint32_t flush_type);
104 	/* flush the vm tlb via pasid */
105 	int (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid,
106 					uint32_t flush_type, bool all_hub);
107 	/* flush the vm tlb via ring */
108 	uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid,
109 				       uint64_t pd_addr);
110 	/* Change the VMID -> PASID mapping */
111 	void (*emit_pasid_mapping)(struct amdgpu_ring *ring, unsigned vmid,
112 				   unsigned pasid);
113 	/* enable/disable PRT support */
114 	void (*set_prt)(struct amdgpu_device *adev, bool enable);
115 	/* map mtype to hardware flags */
116 	uint64_t (*map_mtype)(struct amdgpu_device *adev, uint32_t flags);
117 	/* get the pde for a given mc addr */
118 	void (*get_vm_pde)(struct amdgpu_device *adev, int level,
119 			   u64 *dst, u64 *flags);
120 	/* get the pte flags to use for a BO VA mapping */
121 	void (*get_vm_pte)(struct amdgpu_device *adev,
122 			   struct amdgpu_bo_va_mapping *mapping,
123 			   uint64_t *flags);
124 };
125 
126 struct amdgpu_xgmi {
127 	/* from psp */
128 	u64 node_id;
129 	u64 hive_id;
130 	/* fixed per family */
131 	u64 node_segment_size;
132 	/* physical node (0-3) */
133 	unsigned physical_node_id;
134 	/* number of nodes (0-4) */
135 	unsigned num_physical_nodes;
136 	/* gpu list in the same hive */
137 	struct list_head head;
138 	bool supported;
139 	struct ras_common_if *ras_if;
140 };
141 
142 struct amdgpu_gmc {
143 	/* FB's physical address in MMIO space (for CPU to
144 	 * map FB). This is different compared to the agp/
145 	 * gart/vram_start/end field as the later is from
146 	 * GPU's view and aper_base is from CPU's view.
147 	 */
148 	resource_size_t		aper_size;
149 	resource_size_t		aper_base;
150 	/* for some chips with <= 32MB we need to lie
151 	 * about vram size near mc fb location */
152 	u64			mc_vram_size;
153 	u64			visible_vram_size;
154 	/* AGP aperture start and end in MC address space
155 	 * Driver find a hole in the MC address space
156 	 * to place AGP by setting MC_VM_AGP_BOT/TOP registers
157 	 * Under VMID0, logical address == MC address. AGP
158 	 * aperture maps to physical bus or IOVA addressed.
159 	 * AGP aperture is used to simulate FB in ZFB case.
160 	 * AGP aperture is also used for page table in system
161 	 * memory (mainly for APU).
162 	 *
163 	 */
164 	u64			agp_size;
165 	u64			agp_start;
166 	u64			agp_end;
167 	/* GART aperture start and end in MC address space
168 	 * Driver find a hole in the MC address space
169 	 * to place GART by setting VM_CONTEXT0_PAGE_TABLE_START/END_ADDR
170 	 * registers
171 	 * Under VMID0, logical address inside GART aperture will
172 	 * be translated through gpuvm gart page table to access
173 	 * paged system memory
174 	 */
175 	u64			gart_size;
176 	u64			gart_start;
177 	u64			gart_end;
178 	/* Frame buffer aperture of this GPU device. Different from
179 	 * fb_start (see below), this only covers the local GPU device.
180 	 * Driver get fb_start from MC_VM_FB_LOCATION_BASE (set by vbios)
181 	 * and calculate vram_start of this local device by adding an
182 	 * offset inside the XGMI hive.
183 	 * Under VMID0, logical address == MC address
184 	 */
185 	u64			vram_start;
186 	u64			vram_end;
187 	/* FB region , it's same as local vram region in single GPU, in XGMI
188 	 * configuration, this region covers all GPUs in the same hive ,
189 	 * each GPU in the hive has the same view of this FB region .
190 	 * GPU0's vram starts at offset (0 * segment size) ,
191 	 * GPU1 starts at offset (1 * segment size), etc.
192 	 */
193 	u64			fb_start;
194 	u64			fb_end;
195 	unsigned		vram_width;
196 	u64			real_vram_size;
197 	int			vram_mtrr;
198 	u64                     mc_mask;
199 	const struct firmware   *fw;	/* MC firmware */
200 	uint32_t                fw_version;
201 	struct amdgpu_irq_src	vm_fault;
202 	uint32_t		vram_type;
203 	uint8_t			vram_vendor;
204 	uint32_t                srbm_soft_reset;
205 	bool			prt_warning;
206 	uint64_t		stolen_size;
207 	uint32_t		sdpif_register;
208 	/* apertures */
209 	u64			shared_aperture_start;
210 	u64			shared_aperture_end;
211 	u64			private_aperture_start;
212 	u64			private_aperture_end;
213 	/* protects concurrent invalidation */
214 	spinlock_t		invalidate_lock;
215 	bool			translate_further;
216 	struct kfd_vm_fault_info *vm_fault_info;
217 	atomic_t		vm_fault_info_updated;
218 
219 	struct amdgpu_gmc_fault	fault_ring[AMDGPU_GMC_FAULT_RING_SIZE];
220 	struct {
221 		uint64_t	idx:AMDGPU_GMC_FAULT_RING_ORDER;
222 	} fault_hash[AMDGPU_GMC_FAULT_HASH_SIZE];
223 	uint64_t		last_fault:AMDGPU_GMC_FAULT_RING_ORDER;
224 
225 	bool tmz_enabled;
226 
227 	const struct amdgpu_gmc_funcs	*gmc_funcs;
228 
229 	struct amdgpu_xgmi xgmi;
230 	struct amdgpu_irq_src	ecc_irq;
231 };
232 
233 #define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type)))
234 #define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub) \
235 	((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \
236 	((adev), (pasid), (type), (allhub)))
237 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr))
238 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid))
239 #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags))
240 #define amdgpu_gmc_get_vm_pde(adev, level, dst, flags) (adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (dst), (flags))
241 #define amdgpu_gmc_get_vm_pte(adev, mapping, flags) (adev)->gmc.gmc_funcs->get_vm_pte((adev), (mapping), (flags))
242 
243 /**
244  * amdgpu_gmc_vram_full_visible - Check if full VRAM is visible through the BAR
245  *
246  * @adev: amdgpu_device pointer
247  *
248  * Returns:
249  * True if full VRAM is visible through the BAR
250  */
251 static inline bool amdgpu_gmc_vram_full_visible(struct amdgpu_gmc *gmc)
252 {
253 	WARN_ON(gmc->real_vram_size < gmc->visible_vram_size);
254 
255 	return (gmc->real_vram_size == gmc->visible_vram_size);
256 }
257 
258 /**
259  * amdgpu_gmc_sign_extend - sign extend the given gmc address
260  *
261  * @addr: address to extend
262  */
263 static inline uint64_t amdgpu_gmc_sign_extend(uint64_t addr)
264 {
265 	if (addr >= AMDGPU_GMC_HOLE_START)
266 		addr |= AMDGPU_GMC_HOLE_END;
267 
268 	return addr;
269 }
270 
271 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
272 			       uint64_t *addr, uint64_t *flags);
273 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
274 				uint32_t gpu_page_idx, uint64_t addr,
275 				uint64_t flags);
276 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo);
277 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo);
278 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
279 			      u64 base);
280 void amdgpu_gmc_gart_location(struct amdgpu_device *adev,
281 			      struct amdgpu_gmc *mc);
282 void amdgpu_gmc_agp_location(struct amdgpu_device *adev,
283 			     struct amdgpu_gmc *mc);
284 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr,
285 			      uint16_t pasid, uint64_t timestamp);
286 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
287 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
288 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
289 
290 extern void amdgpu_gmc_tmz_set(struct amdgpu_device *adev);
291 
292 #endif
293