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