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 /*
89  * GPU MC structures, functions & helpers
90  */
91 struct amdgpu_gmc_funcs {
92 	/* flush the vm tlb via mmio */
93 	void (*flush_gpu_tlb)(struct amdgpu_device *adev, uint32_t vmid,
94 				uint32_t vmhub, uint32_t flush_type);
95 	/* flush the vm tlb via pasid */
96 	int (*flush_gpu_tlb_pasid)(struct amdgpu_device *adev, uint16_t pasid,
97 					uint32_t flush_type, bool all_hub);
98 	/* flush the vm tlb via ring */
99 	uint64_t (*emit_flush_gpu_tlb)(struct amdgpu_ring *ring, unsigned vmid,
100 				       uint64_t pd_addr);
101 	/* Change the VMID -> PASID mapping */
102 	void (*emit_pasid_mapping)(struct amdgpu_ring *ring, unsigned vmid,
103 				   unsigned pasid);
104 	/* enable/disable PRT support */
105 	void (*set_prt)(struct amdgpu_device *adev, bool enable);
106 	/* map mtype to hardware flags */
107 	uint64_t (*map_mtype)(struct amdgpu_device *adev, uint32_t flags);
108 	/* get the pde for a given mc addr */
109 	void (*get_vm_pde)(struct amdgpu_device *adev, int level,
110 			   u64 *dst, u64 *flags);
111 	/* get the pte flags to use for a BO VA mapping */
112 	void (*get_vm_pte)(struct amdgpu_device *adev,
113 			   struct amdgpu_bo_va_mapping *mapping,
114 			   uint64_t *flags);
115 };
116 
117 struct amdgpu_xgmi {
118 	/* from psp */
119 	u64 node_id;
120 	u64 hive_id;
121 	/* fixed per family */
122 	u64 node_segment_size;
123 	/* physical node (0-3) */
124 	unsigned physical_node_id;
125 	/* number of nodes (0-4) */
126 	unsigned num_physical_nodes;
127 	/* gpu list in the same hive */
128 	struct list_head head;
129 	bool supported;
130 	struct ras_common_if *ras_if;
131 };
132 
133 struct amdgpu_gmc {
134 	/* FB's physical address in MMIO space (for CPU to
135 	 * map FB). This is different compared to the agp/
136 	 * gart/vram_start/end field as the later is from
137 	 * GPU's view and aper_base is from CPU's view.
138 	 */
139 	resource_size_t		aper_size;
140 	resource_size_t		aper_base;
141 	/* for some chips with <= 32MB we need to lie
142 	 * about vram size near mc fb location */
143 	u64			mc_vram_size;
144 	u64			visible_vram_size;
145 	/* AGP aperture start and end in MC address space
146 	 * Driver find a hole in the MC address space
147 	 * to place AGP by setting MC_VM_AGP_BOT/TOP registers
148 	 * Under VMID0, logical address == MC address. AGP
149 	 * aperture maps to physical bus or IOVA addressed.
150 	 * AGP aperture is used to simulate FB in ZFB case.
151 	 * AGP aperture is also used for page table in system
152 	 * memory (mainly for APU).
153 	 *
154 	 */
155 	u64			agp_size;
156 	u64			agp_start;
157 	u64			agp_end;
158 	/* GART aperture start and end in MC address space
159 	 * Driver find a hole in the MC address space
160 	 * to place GART by setting VM_CONTEXT0_PAGE_TABLE_START/END_ADDR
161 	 * registers
162 	 * Under VMID0, logical address inside GART aperture will
163 	 * be translated through gpuvm gart page table to access
164 	 * paged system memory
165 	 */
166 	u64			gart_size;
167 	u64			gart_start;
168 	u64			gart_end;
169 	/* Frame buffer aperture of this GPU device. Different from
170 	 * fb_start (see below), this only covers the local GPU device.
171 	 * Driver get fb_start from MC_VM_FB_LOCATION_BASE (set by vbios)
172 	 * and calculate vram_start of this local device by adding an
173 	 * offset inside the XGMI hive.
174 	 * Under VMID0, logical address == MC address
175 	 */
176 	u64			vram_start;
177 	u64			vram_end;
178 	/* FB region , it's same as local vram region in single GPU, in XGMI
179 	 * configuration, this region covers all GPUs in the same hive ,
180 	 * each GPU in the hive has the same view of this FB region .
181 	 * GPU0's vram starts at offset (0 * segment size) ,
182 	 * GPU1 starts at offset (1 * segment size), etc.
183 	 */
184 	u64			fb_start;
185 	u64			fb_end;
186 	unsigned		vram_width;
187 	u64			real_vram_size;
188 	int			vram_mtrr;
189 	u64                     mc_mask;
190 	const struct firmware   *fw;	/* MC firmware */
191 	uint32_t                fw_version;
192 	struct amdgpu_irq_src	vm_fault;
193 	uint32_t		vram_type;
194 	uint8_t			vram_vendor;
195 	uint32_t                srbm_soft_reset;
196 	bool			prt_warning;
197 	uint64_t		stolen_size;
198 	/* apertures */
199 	u64			shared_aperture_start;
200 	u64			shared_aperture_end;
201 	u64			private_aperture_start;
202 	u64			private_aperture_end;
203 	/* protects concurrent invalidation */
204 	spinlock_t		invalidate_lock;
205 	bool			translate_further;
206 	struct kfd_vm_fault_info *vm_fault_info;
207 	atomic_t		vm_fault_info_updated;
208 
209 	struct amdgpu_gmc_fault	fault_ring[AMDGPU_GMC_FAULT_RING_SIZE];
210 	struct {
211 		uint64_t	idx:AMDGPU_GMC_FAULT_RING_ORDER;
212 	} fault_hash[AMDGPU_GMC_FAULT_HASH_SIZE];
213 	uint64_t		last_fault:AMDGPU_GMC_FAULT_RING_ORDER;
214 
215 	const struct amdgpu_gmc_funcs	*gmc_funcs;
216 
217 	struct amdgpu_xgmi xgmi;
218 	struct amdgpu_irq_src	ecc_irq;
219 };
220 
221 #define amdgpu_gmc_flush_gpu_tlb(adev, vmid, vmhub, type) ((adev)->gmc.gmc_funcs->flush_gpu_tlb((adev), (vmid), (vmhub), (type)))
222 #define amdgpu_gmc_flush_gpu_tlb_pasid(adev, pasid, type, allhub) \
223 	((adev)->gmc.gmc_funcs->flush_gpu_tlb_pasid \
224 	((adev), (pasid), (type), (allhub)))
225 #define amdgpu_gmc_emit_flush_gpu_tlb(r, vmid, addr) (r)->adev->gmc.gmc_funcs->emit_flush_gpu_tlb((r), (vmid), (addr))
226 #define amdgpu_gmc_emit_pasid_mapping(r, vmid, pasid) (r)->adev->gmc.gmc_funcs->emit_pasid_mapping((r), (vmid), (pasid))
227 #define amdgpu_gmc_map_mtype(adev, flags) (adev)->gmc.gmc_funcs->map_mtype((adev),(flags))
228 #define amdgpu_gmc_get_vm_pde(adev, level, dst, flags) (adev)->gmc.gmc_funcs->get_vm_pde((adev), (level), (dst), (flags))
229 #define amdgpu_gmc_get_vm_pte(adev, mapping, flags) (adev)->gmc.gmc_funcs->get_vm_pte((adev), (mapping), (flags))
230 
231 /**
232  * amdgpu_gmc_vram_full_visible - Check if full VRAM is visible through the BAR
233  *
234  * @adev: amdgpu_device pointer
235  *
236  * Returns:
237  * True if full VRAM is visible through the BAR
238  */
239 static inline bool amdgpu_gmc_vram_full_visible(struct amdgpu_gmc *gmc)
240 {
241 	WARN_ON(gmc->real_vram_size < gmc->visible_vram_size);
242 
243 	return (gmc->real_vram_size == gmc->visible_vram_size);
244 }
245 
246 /**
247  * amdgpu_gmc_sign_extend - sign extend the given gmc address
248  *
249  * @addr: address to extend
250  */
251 static inline uint64_t amdgpu_gmc_sign_extend(uint64_t addr)
252 {
253 	if (addr >= AMDGPU_GMC_HOLE_START)
254 		addr |= AMDGPU_GMC_HOLE_END;
255 
256 	return addr;
257 }
258 
259 void amdgpu_gmc_get_pde_for_bo(struct amdgpu_bo *bo, int level,
260 			       uint64_t *addr, uint64_t *flags);
261 int amdgpu_gmc_set_pte_pde(struct amdgpu_device *adev, void *cpu_pt_addr,
262 				uint32_t gpu_page_idx, uint64_t addr,
263 				uint64_t flags);
264 uint64_t amdgpu_gmc_pd_addr(struct amdgpu_bo *bo);
265 uint64_t amdgpu_gmc_agp_addr(struct ttm_buffer_object *bo);
266 void amdgpu_gmc_vram_location(struct amdgpu_device *adev, struct amdgpu_gmc *mc,
267 			      u64 base);
268 void amdgpu_gmc_gart_location(struct amdgpu_device *adev,
269 			      struct amdgpu_gmc *mc);
270 void amdgpu_gmc_agp_location(struct amdgpu_device *adev,
271 			     struct amdgpu_gmc *mc);
272 bool amdgpu_gmc_filter_faults(struct amdgpu_device *adev, uint64_t addr,
273 			      uint16_t pasid, uint64_t timestamp);
274 int amdgpu_gmc_ras_late_init(struct amdgpu_device *adev);
275 void amdgpu_gmc_ras_fini(struct amdgpu_device *adev);
276 int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev);
277 
278 #endif
279