xref: /openbmc/linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h (revision 61c1f340bc809a1ca1e3c8794207a91cde1a7c78)
1 /*
2  * Copyright 2019 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  *
22  */
23 
24 #ifndef __AMDGPU_MES_H__
25 #define __AMDGPU_MES_H__
26 
27 #include "amdgpu_irq.h"
28 #include "kgd_kfd_interface.h"
29 #include "amdgpu_gfx.h"
30 #include <linux/sched/mm.h>
31 
32 #define AMDGPU_MES_MAX_COMPUTE_PIPES        8
33 #define AMDGPU_MES_MAX_GFX_PIPES            2
34 #define AMDGPU_MES_MAX_SDMA_PIPES           2
35 
36 #define AMDGPU_MES_API_VERSION_SHIFT	12
37 #define AMDGPU_MES_FEAT_VERSION_SHIFT	24
38 
39 #define AMDGPU_MES_VERSION_MASK		0x00000fff
40 #define AMDGPU_MES_API_VERSION_MASK	0x00fff000
41 #define AMDGPU_MES_FEAT_VERSION_MASK	0xff000000
42 
43 enum amdgpu_mes_priority_level {
44 	AMDGPU_MES_PRIORITY_LEVEL_LOW       = 0,
45 	AMDGPU_MES_PRIORITY_LEVEL_NORMAL    = 1,
46 	AMDGPU_MES_PRIORITY_LEVEL_MEDIUM    = 2,
47 	AMDGPU_MES_PRIORITY_LEVEL_HIGH      = 3,
48 	AMDGPU_MES_PRIORITY_LEVEL_REALTIME  = 4,
49 	AMDGPU_MES_PRIORITY_NUM_LEVELS
50 };
51 
52 #define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */
53 #define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */
54 
55 struct amdgpu_mes_funcs;
56 
57 enum admgpu_mes_pipe {
58 	AMDGPU_MES_SCHED_PIPE = 0,
59 	AMDGPU_MES_KIQ_PIPE,
60 	AMDGPU_MAX_MES_PIPES = 2,
61 };
62 
63 struct amdgpu_mes {
64 	struct amdgpu_device            *adev;
65 
66 	struct mutex                    mutex_hidden;
67 
68 	struct idr                      pasid_idr;
69 	struct idr                      gang_id_idr;
70 	struct idr                      queue_id_idr;
71 	struct ida                      doorbell_ida;
72 
73 	spinlock_t                      queue_id_lock;
74 
75 	uint32_t			sched_version;
76 	uint32_t			kiq_version;
77 
78 	uint32_t                        total_max_queue;
79 	uint32_t                        doorbell_id_offset;
80 	uint32_t                        max_doorbell_slices;
81 
82 	uint64_t                        default_process_quantum;
83 	uint64_t                        default_gang_quantum;
84 
85 	struct amdgpu_ring              ring;
86 
87 	const struct firmware           *fw[AMDGPU_MAX_MES_PIPES];
88 
89 	/* mes ucode */
90 	struct amdgpu_bo		*ucode_fw_obj[AMDGPU_MAX_MES_PIPES];
91 	uint64_t			ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
92 	uint32_t			*ucode_fw_ptr[AMDGPU_MAX_MES_PIPES];
93 	uint32_t                        ucode_fw_version[AMDGPU_MAX_MES_PIPES];
94 	uint64_t                        uc_start_addr[AMDGPU_MAX_MES_PIPES];
95 
96 	/* mes ucode data */
97 	struct amdgpu_bo		*data_fw_obj[AMDGPU_MAX_MES_PIPES];
98 	uint64_t			data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
99 	uint32_t			*data_fw_ptr[AMDGPU_MAX_MES_PIPES];
100 	uint32_t                        data_fw_version[AMDGPU_MAX_MES_PIPES];
101 	uint64_t                        data_start_addr[AMDGPU_MAX_MES_PIPES];
102 
103 	/* eop gpu obj */
104 	struct amdgpu_bo		*eop_gpu_obj[AMDGPU_MAX_MES_PIPES];
105 	uint64_t                        eop_gpu_addr[AMDGPU_MAX_MES_PIPES];
106 
107 	void                            *mqd_backup[AMDGPU_MAX_MES_PIPES];
108 	struct amdgpu_irq_src	        irq[AMDGPU_MAX_MES_PIPES];
109 
110 	uint32_t                        vmid_mask_gfxhub;
111 	uint32_t                        vmid_mask_mmhub;
112 	uint32_t                        compute_hqd_mask[AMDGPU_MES_MAX_COMPUTE_PIPES];
113 	uint32_t                        gfx_hqd_mask[AMDGPU_MES_MAX_GFX_PIPES];
114 	uint32_t                        sdma_hqd_mask[AMDGPU_MES_MAX_SDMA_PIPES];
115 	uint32_t                        agreegated_doorbells[AMDGPU_MES_PRIORITY_NUM_LEVELS];
116 	uint32_t                        sch_ctx_offs;
117 	uint64_t			sch_ctx_gpu_addr;
118 	uint64_t			*sch_ctx_ptr;
119 	uint32_t			query_status_fence_offs;
120 	uint64_t			query_status_fence_gpu_addr;
121 	uint64_t			*query_status_fence_ptr;
122 	uint32_t                        read_val_offs;
123 	uint64_t			read_val_gpu_addr;
124 	uint32_t			*read_val_ptr;
125 
126 	uint32_t			saved_flags;
127 
128 	/* initialize kiq pipe */
129 	int                             (*kiq_hw_init)(struct amdgpu_device *adev);
130 	int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
131 
132 	/* ip specific functions */
133 	const struct amdgpu_mes_funcs   *funcs;
134 };
135 
136 struct amdgpu_mes_process {
137 	int			pasid;
138 	struct			amdgpu_vm *vm;
139 	uint64_t		pd_gpu_addr;
140 	struct amdgpu_bo 	*proc_ctx_bo;
141 	uint64_t 		proc_ctx_gpu_addr;
142 	void 			*proc_ctx_cpu_ptr;
143 	uint64_t 		process_quantum;
144 	struct 			list_head gang_list;
145 	uint32_t 		doorbell_index;
146 	unsigned long 		*doorbell_bitmap;
147 	struct mutex		doorbell_lock;
148 };
149 
150 struct amdgpu_mes_gang {
151 	int 				gang_id;
152 	int 				priority;
153 	int 				inprocess_gang_priority;
154 	int 				global_priority_level;
155 	struct list_head 		list;
156 	struct amdgpu_mes_process 	*process;
157 	struct amdgpu_bo 		*gang_ctx_bo;
158 	uint64_t 			gang_ctx_gpu_addr;
159 	void 				*gang_ctx_cpu_ptr;
160 	uint64_t 			gang_quantum;
161 	struct list_head 		queue_list;
162 };
163 
164 struct amdgpu_mes_queue {
165 	struct list_head 		list;
166 	struct amdgpu_mes_gang 		*gang;
167 	int 				queue_id;
168 	uint64_t 			doorbell_off;
169 	struct amdgpu_bo		*mqd_obj;
170 	void				*mqd_cpu_ptr;
171 	uint64_t 			mqd_gpu_addr;
172 	uint64_t 			wptr_gpu_addr;
173 	int 				queue_type;
174 	int 				paging;
175 	struct amdgpu_ring 		*ring;
176 };
177 
178 struct amdgpu_mes_queue_properties {
179 	int 			queue_type;
180 	uint64_t                hqd_base_gpu_addr;
181 	uint64_t                rptr_gpu_addr;
182 	uint64_t                wptr_gpu_addr;
183 	uint64_t                wptr_mc_addr;
184 	uint32_t                queue_size;
185 	uint64_t                eop_gpu_addr;
186 	uint32_t                hqd_pipe_priority;
187 	uint32_t                hqd_queue_priority;
188 	bool 			paging;
189 	struct amdgpu_ring 	*ring;
190 	/* out */
191 	uint64_t       		doorbell_off;
192 };
193 
194 struct amdgpu_mes_gang_properties {
195 	uint32_t 	priority;
196 	uint32_t 	gang_quantum;
197 	uint32_t 	inprocess_gang_priority;
198 	uint32_t 	priority_level;
199 	int 		global_priority_level;
200 };
201 
202 struct mes_add_queue_input {
203 	uint32_t	process_id;
204 	uint64_t	page_table_base_addr;
205 	uint64_t	process_va_start;
206 	uint64_t	process_va_end;
207 	uint64_t	process_quantum;
208 	uint64_t	process_context_addr;
209 	uint64_t	gang_quantum;
210 	uint64_t	gang_context_addr;
211 	uint32_t	inprocess_gang_priority;
212 	uint32_t	gang_global_priority_level;
213 	uint32_t	doorbell_offset;
214 	uint64_t	mqd_addr;
215 	uint64_t	wptr_addr;
216 	uint64_t	wptr_mc_addr;
217 	uint32_t	queue_type;
218 	uint32_t	paging;
219 	uint32_t        gws_base;
220 	uint32_t        gws_size;
221 	uint64_t	tba_addr;
222 	uint64_t	tma_addr;
223 	uint32_t	is_kfd_process;
224 };
225 
226 struct mes_remove_queue_input {
227 	uint32_t	doorbell_offset;
228 	uint64_t	gang_context_addr;
229 };
230 
231 struct mes_unmap_legacy_queue_input {
232 	enum amdgpu_unmap_queues_action    action;
233 	uint32_t                           queue_type;
234 	uint32_t                           doorbell_offset;
235 	uint32_t                           pipe_id;
236 	uint32_t                           queue_id;
237 	uint64_t                           trail_fence_addr;
238 	uint64_t                           trail_fence_data;
239 };
240 
241 struct mes_suspend_gang_input {
242 	bool		suspend_all_gangs;
243 	uint64_t	gang_context_addr;
244 	uint64_t	suspend_fence_addr;
245 	uint32_t	suspend_fence_value;
246 };
247 
248 struct mes_resume_gang_input {
249 	bool		resume_all_gangs;
250 	uint64_t	gang_context_addr;
251 };
252 
253 enum mes_misc_opcode {
254 	MES_MISC_OP_WRITE_REG,
255 	MES_MISC_OP_READ_REG,
256 	MES_MISC_OP_WRM_REG_WAIT,
257 	MES_MISC_OP_WRM_REG_WR_WAIT,
258 };
259 
260 struct mes_misc_op_input {
261 	enum mes_misc_opcode op;
262 
263 	union {
264 		struct {
265 			uint32_t                  reg_offset;
266 			uint64_t                  buffer_addr;
267 		} read_reg;
268 
269 		struct {
270 			uint32_t                  reg_offset;
271 			uint32_t                  reg_value;
272 		} write_reg;
273 
274 		struct {
275 			uint32_t                   ref;
276 			uint32_t                   mask;
277 			uint32_t                   reg0;
278 			uint32_t                   reg1;
279 		} wrm_reg;
280 	};
281 };
282 
283 struct amdgpu_mes_funcs {
284 	int (*add_hw_queue)(struct amdgpu_mes *mes,
285 			    struct mes_add_queue_input *input);
286 
287 	int (*remove_hw_queue)(struct amdgpu_mes *mes,
288 			       struct mes_remove_queue_input *input);
289 
290 	int (*unmap_legacy_queue)(struct amdgpu_mes *mes,
291 				  struct mes_unmap_legacy_queue_input *input);
292 
293 	int (*suspend_gang)(struct amdgpu_mes *mes,
294 			    struct mes_suspend_gang_input *input);
295 
296 	int (*resume_gang)(struct amdgpu_mes *mes,
297 			   struct mes_resume_gang_input *input);
298 
299 	int (*misc_op)(struct amdgpu_mes *mes,
300 		       struct mes_misc_op_input *input);
301 };
302 
303 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
304 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev))
305 
306 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs);
307 
308 int amdgpu_mes_init(struct amdgpu_device *adev);
309 void amdgpu_mes_fini(struct amdgpu_device *adev);
310 
311 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
312 			      struct amdgpu_vm *vm);
313 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid);
314 
315 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
316 			struct amdgpu_mes_gang_properties *gprops,
317 			int *gang_id);
318 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id);
319 
320 int amdgpu_mes_suspend(struct amdgpu_device *adev);
321 int amdgpu_mes_resume(struct amdgpu_device *adev);
322 
323 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
324 			    struct amdgpu_mes_queue_properties *qprops,
325 			    int *queue_id);
326 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id);
327 
328 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
329 				  struct amdgpu_ring *ring,
330 				  enum amdgpu_unmap_queues_action action,
331 				  u64 gpu_addr, u64 seq);
332 
333 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
334 int amdgpu_mes_wreg(struct amdgpu_device *adev,
335 		    uint32_t reg, uint32_t val);
336 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
337 			uint32_t val, uint32_t mask);
338 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
339 				  uint32_t reg0, uint32_t reg1,
340 				  uint32_t ref, uint32_t mask);
341 
342 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
343 			int queue_type, int idx,
344 			struct amdgpu_mes_ctx_data *ctx_data,
345 			struct amdgpu_ring **out);
346 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
347 			    struct amdgpu_ring *ring);
348 
349 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
350 				   struct amdgpu_mes_ctx_data *ctx_data);
351 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data);
352 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
353 				 struct amdgpu_vm *vm,
354 				 struct amdgpu_mes_ctx_data *ctx_data);
355 
356 int amdgpu_mes_self_test(struct amdgpu_device *adev);
357 
358 int amdgpu_mes_alloc_process_doorbells(struct amdgpu_device *adev,
359 					unsigned int *doorbell_index);
360 void amdgpu_mes_free_process_doorbells(struct amdgpu_device *adev,
361 					unsigned int doorbell_index);
362 unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar(
363 					struct amdgpu_device *adev,
364 					uint32_t doorbell_index,
365 					unsigned int doorbell_id);
366 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev);
367 
368 /*
369  * MES lock can be taken in MMU notifiers.
370  *
371  * A bit more detail about why to set no-FS reclaim with MES lock:
372  *
373  * The purpose of the MMU notifier is to stop GPU access to memory so
374  * that the Linux VM subsystem can move pages around safely. This is
375  * done by preempting user mode queues for the affected process. When
376  * MES is used, MES lock needs to be taken to preempt the queues.
377  *
378  * The MMU notifier callback entry point in the driver is
379  * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from
380  * there is:
381  * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm ->
382  * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues
383  *
384  * The last part of the chain is a function pointer where we take the
385  * MES lock.
386  *
387  * The problem with taking locks in the MMU notifier is, that MMU
388  * notifiers can be called in reclaim-FS context. That's where the
389  * kernel frees up pages to make room for new page allocations under
390  * memory pressure. While we are running in reclaim-FS context, we must
391  * not trigger another memory reclaim operation because that would
392  * recursively reenter the reclaim code and cause a deadlock. The
393  * memalloc_nofs_save/restore calls guarantee that.
394  *
395  * In addition we also need to avoid lock dependencies on other locks taken
396  * under the MES lock, for example reservation locks. Here is a possible
397  * scenario of a deadlock:
398  * Thread A: takes and holds reservation lock | triggers reclaim-FS |
399  * MMU notifier | blocks trying to take MES lock
400  * Thread B: takes and holds MES lock | blocks trying to take reservation lock
401  *
402  * In this scenario Thread B gets involved in a deadlock even without
403  * triggering a reclaim-FS operation itself.
404  * To fix this and break the lock dependency chain you'd need to either:
405  * 1. protect reservation locks with memalloc_nofs_save/restore, or
406  * 2. avoid taking reservation locks under the MES lock.
407  *
408  * Reservation locks are taken all over the kernel in different subsystems, we
409  * have no control over them and their lock dependencies.So the only workable
410  * solution is to avoid taking other locks under the MES lock.
411  * As a result, make sure no reclaim-FS happens while holding this lock anywhere
412  * to prevent deadlocks when an MMU notifier runs in reclaim-FS context.
413  */
414 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes)
415 {
416 	mutex_lock(&mes->mutex_hidden);
417 	mes->saved_flags = memalloc_noreclaim_save();
418 }
419 
420 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes)
421 {
422 	memalloc_noreclaim_restore(mes->saved_flags);
423 	mutex_unlock(&mes->mutex_hidden);
424 }
425 #endif /* __AMDGPU_MES_H__ */
426