xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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 "amdgpu_doorbell.h"
31 #include <linux/sched/mm.h>
32 
33 #define AMDGPU_MES_MAX_COMPUTE_PIPES        8
34 #define AMDGPU_MES_MAX_GFX_PIPES            2
35 #define AMDGPU_MES_MAX_SDMA_PIPES           2
36 
37 #define AMDGPU_MES_API_VERSION_SHIFT	12
38 #define AMDGPU_MES_FEAT_VERSION_SHIFT	24
39 
40 #define AMDGPU_MES_VERSION_MASK		0x00000fff
41 #define AMDGPU_MES_API_VERSION_MASK	0x00fff000
42 #define AMDGPU_MES_FEAT_VERSION_MASK	0xff000000
43 #define AMDGPU_MES_MSCRATCH_SIZE	0x8000
44 
45 enum amdgpu_mes_priority_level {
46 	AMDGPU_MES_PRIORITY_LEVEL_LOW       = 0,
47 	AMDGPU_MES_PRIORITY_LEVEL_NORMAL    = 1,
48 	AMDGPU_MES_PRIORITY_LEVEL_MEDIUM    = 2,
49 	AMDGPU_MES_PRIORITY_LEVEL_HIGH      = 3,
50 	AMDGPU_MES_PRIORITY_LEVEL_REALTIME  = 4,
51 	AMDGPU_MES_PRIORITY_NUM_LEVELS
52 };
53 
54 #define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */
55 #define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */
56 
57 struct amdgpu_mes_funcs;
58 
59 enum admgpu_mes_pipe {
60 	AMDGPU_MES_SCHED_PIPE = 0,
61 	AMDGPU_MES_KIQ_PIPE,
62 	AMDGPU_MAX_MES_PIPES = 2,
63 };
64 
65 struct amdgpu_mes {
66 	struct amdgpu_device            *adev;
67 
68 	struct mutex                    mutex_hidden;
69 
70 	struct idr                      pasid_idr;
71 	struct idr                      gang_id_idr;
72 	struct idr                      queue_id_idr;
73 	struct ida                      doorbell_ida;
74 
75 	spinlock_t                      queue_id_lock;
76 
77 	uint32_t			sched_version;
78 	uint32_t			kiq_version;
79 	uint32_t			fw_version[AMDGPU_MAX_MES_PIPES];
80 	bool                            enable_legacy_queue_map;
81 
82 	uint32_t                        total_max_queue;
83 	uint32_t                        max_doorbell_slices;
84 
85 	uint64_t                        default_process_quantum;
86 	uint64_t                        default_gang_quantum;
87 
88 	struct amdgpu_ring              ring[AMDGPU_MAX_MES_PIPES];
89 	spinlock_t                      ring_lock[AMDGPU_MAX_MES_PIPES];
90 
91 	const struct firmware           *fw[AMDGPU_MAX_MES_PIPES];
92 
93 	/* mes ucode */
94 	struct amdgpu_bo		*ucode_fw_obj[AMDGPU_MAX_MES_PIPES];
95 	uint64_t			ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
96 	uint32_t			*ucode_fw_ptr[AMDGPU_MAX_MES_PIPES];
97 	uint64_t                        uc_start_addr[AMDGPU_MAX_MES_PIPES];
98 
99 	/* mes ucode data */
100 	struct amdgpu_bo		*data_fw_obj[AMDGPU_MAX_MES_PIPES];
101 	uint64_t			data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
102 	uint32_t			*data_fw_ptr[AMDGPU_MAX_MES_PIPES];
103 	uint64_t                        data_start_addr[AMDGPU_MAX_MES_PIPES];
104 
105 	/* eop gpu obj */
106 	struct amdgpu_bo		*eop_gpu_obj[AMDGPU_MAX_MES_PIPES];
107 	uint64_t                        eop_gpu_addr[AMDGPU_MAX_MES_PIPES];
108 
109 	void                            *mqd_backup[AMDGPU_MAX_MES_PIPES];
110 	struct amdgpu_irq_src	        irq[AMDGPU_MAX_MES_PIPES];
111 
112 	uint32_t                        vmid_mask_gfxhub;
113 	uint32_t                        vmid_mask_mmhub;
114 	uint32_t                        compute_hqd_mask[AMDGPU_MES_MAX_COMPUTE_PIPES];
115 	uint32_t                        gfx_hqd_mask[AMDGPU_MES_MAX_GFX_PIPES];
116 	uint32_t                        sdma_hqd_mask[AMDGPU_MES_MAX_SDMA_PIPES];
117 	uint32_t                        aggregated_doorbells[AMDGPU_MES_PRIORITY_NUM_LEVELS];
118 	uint32_t                        sch_ctx_offs[AMDGPU_MAX_MES_PIPES];
119 	uint64_t			sch_ctx_gpu_addr[AMDGPU_MAX_MES_PIPES];
120 	uint64_t			*sch_ctx_ptr[AMDGPU_MAX_MES_PIPES];
121 	uint32_t			query_status_fence_offs[AMDGPU_MAX_MES_PIPES];
122 	uint64_t			query_status_fence_gpu_addr[AMDGPU_MAX_MES_PIPES];
123 	uint64_t			*query_status_fence_ptr[AMDGPU_MAX_MES_PIPES];
124 
125 	uint32_t			saved_flags;
126 
127 	/* initialize kiq pipe */
128 	int                             (*kiq_hw_init)(struct amdgpu_device *adev);
129 	int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
130 
131 	/* MES doorbells */
132 	uint32_t			db_start_dw_offset;
133 	uint32_t			num_mes_dbs;
134 	unsigned long			*doorbell_bitmap;
135 
136 	/* MES event log buffer */
137 	uint32_t			event_log_size;
138 	struct amdgpu_bo	*event_log_gpu_obj;
139 	uint64_t			event_log_gpu_addr;
140 	void				*event_log_cpu_addr;
141 
142 	/* ip specific functions */
143 	const struct amdgpu_mes_funcs   *funcs;
144 
145 	/* mes resource_1 bo*/
146 	struct amdgpu_bo    *resource_1;
147 	uint64_t            resource_1_gpu_addr;
148 	void                *resource_1_addr;
149 
150 };
151 
152 struct amdgpu_mes_process {
153 	int			pasid;
154 	struct			amdgpu_vm *vm;
155 	uint64_t		pd_gpu_addr;
156 	struct amdgpu_bo 	*proc_ctx_bo;
157 	uint64_t 		proc_ctx_gpu_addr;
158 	void 			*proc_ctx_cpu_ptr;
159 	uint64_t 		process_quantum;
160 	struct 			list_head gang_list;
161 	uint32_t 		doorbell_index;
162 	struct mutex		doorbell_lock;
163 };
164 
165 struct amdgpu_mes_gang {
166 	int 				gang_id;
167 	int 				priority;
168 	int 				inprocess_gang_priority;
169 	int 				global_priority_level;
170 	struct list_head 		list;
171 	struct amdgpu_mes_process 	*process;
172 	struct amdgpu_bo 		*gang_ctx_bo;
173 	uint64_t 			gang_ctx_gpu_addr;
174 	void 				*gang_ctx_cpu_ptr;
175 	uint64_t 			gang_quantum;
176 	struct list_head 		queue_list;
177 };
178 
179 struct amdgpu_mes_queue {
180 	struct list_head 		list;
181 	struct amdgpu_mes_gang 		*gang;
182 	int 				queue_id;
183 	uint64_t 			doorbell_off;
184 	struct amdgpu_bo		*mqd_obj;
185 	void				*mqd_cpu_ptr;
186 	uint64_t 			mqd_gpu_addr;
187 	uint64_t 			wptr_gpu_addr;
188 	int 				queue_type;
189 	int 				paging;
190 	struct amdgpu_ring 		*ring;
191 };
192 
193 struct amdgpu_mes_queue_properties {
194 	int 			queue_type;
195 	uint64_t                hqd_base_gpu_addr;
196 	uint64_t                rptr_gpu_addr;
197 	uint64_t                wptr_gpu_addr;
198 	uint64_t                wptr_mc_addr;
199 	uint32_t                queue_size;
200 	uint64_t                eop_gpu_addr;
201 	uint32_t                hqd_pipe_priority;
202 	uint32_t                hqd_queue_priority;
203 	bool 			paging;
204 	struct amdgpu_ring 	*ring;
205 	/* out */
206 	uint64_t       		doorbell_off;
207 };
208 
209 struct amdgpu_mes_gang_properties {
210 	uint32_t 	priority;
211 	uint32_t 	gang_quantum;
212 	uint32_t 	inprocess_gang_priority;
213 	uint32_t 	priority_level;
214 	int 		global_priority_level;
215 };
216 
217 struct mes_add_queue_input {
218 	uint32_t	process_id;
219 	uint64_t	page_table_base_addr;
220 	uint64_t	process_va_start;
221 	uint64_t	process_va_end;
222 	uint64_t	process_quantum;
223 	uint64_t	process_context_addr;
224 	uint64_t	gang_quantum;
225 	uint64_t	gang_context_addr;
226 	uint32_t	inprocess_gang_priority;
227 	uint32_t	gang_global_priority_level;
228 	uint32_t	doorbell_offset;
229 	uint64_t	mqd_addr;
230 	uint64_t	wptr_addr;
231 	uint64_t	wptr_mc_addr;
232 	uint32_t	queue_type;
233 	uint32_t	paging;
234 	uint32_t        gws_base;
235 	uint32_t        gws_size;
236 	uint64_t	tba_addr;
237 	uint64_t	tma_addr;
238 	uint32_t	trap_en;
239 	uint32_t	skip_process_ctx_clear;
240 	uint32_t	is_kfd_process;
241 	uint32_t	is_aql_queue;
242 	uint32_t	queue_size;
243 	uint32_t	exclusively_scheduled;
244 };
245 
246 struct mes_remove_queue_input {
247 	uint32_t	doorbell_offset;
248 	uint64_t	gang_context_addr;
249 };
250 
251 struct mes_reset_queue_input {
252 	uint32_t	doorbell_offset;
253 	uint64_t	gang_context_addr;
254 	bool		use_mmio;
255 	uint32_t	queue_type;
256 	uint32_t	me_id;
257 	uint32_t	pipe_id;
258 	uint32_t	queue_id;
259 	uint32_t	xcc_id;
260 	uint32_t	vmid;
261 };
262 
263 struct mes_map_legacy_queue_input {
264 	uint32_t                           queue_type;
265 	uint32_t                           doorbell_offset;
266 	uint32_t                           pipe_id;
267 	uint32_t                           queue_id;
268 	uint64_t                           mqd_addr;
269 	uint64_t                           wptr_addr;
270 };
271 
272 struct mes_unmap_legacy_queue_input {
273 	enum amdgpu_unmap_queues_action    action;
274 	uint32_t                           queue_type;
275 	uint32_t                           doorbell_offset;
276 	uint32_t                           pipe_id;
277 	uint32_t                           queue_id;
278 	uint64_t                           trail_fence_addr;
279 	uint64_t                           trail_fence_data;
280 };
281 
282 struct mes_suspend_gang_input {
283 	bool		suspend_all_gangs;
284 	uint64_t	gang_context_addr;
285 	uint64_t	suspend_fence_addr;
286 	uint32_t	suspend_fence_value;
287 };
288 
289 struct mes_resume_gang_input {
290 	bool		resume_all_gangs;
291 	uint64_t	gang_context_addr;
292 };
293 
294 struct mes_reset_legacy_queue_input {
295 	uint32_t                           queue_type;
296 	uint32_t                           doorbell_offset;
297 	bool                               use_mmio;
298 	uint32_t                           me_id;
299 	uint32_t                           pipe_id;
300 	uint32_t                           queue_id;
301 	uint64_t                           mqd_addr;
302 	uint64_t                           wptr_addr;
303 	uint32_t                           vmid;
304 };
305 
306 enum mes_misc_opcode {
307 	MES_MISC_OP_WRITE_REG,
308 	MES_MISC_OP_READ_REG,
309 	MES_MISC_OP_WRM_REG_WAIT,
310 	MES_MISC_OP_WRM_REG_WR_WAIT,
311 	MES_MISC_OP_SET_SHADER_DEBUGGER,
312 	MES_MISC_OP_CHANGE_CONFIG,
313 };
314 
315 struct mes_misc_op_input {
316 	enum mes_misc_opcode op;
317 
318 	union {
319 		struct {
320 			uint32_t                  reg_offset;
321 			uint64_t                  buffer_addr;
322 		} read_reg;
323 
324 		struct {
325 			uint32_t                  reg_offset;
326 			uint32_t                  reg_value;
327 		} write_reg;
328 
329 		struct {
330 			uint32_t                   ref;
331 			uint32_t                   mask;
332 			uint32_t                   reg0;
333 			uint32_t                   reg1;
334 		} wrm_reg;
335 
336 		struct {
337 			uint64_t process_context_addr;
338 			union {
339 				struct {
340 					uint32_t single_memop : 1;
341 					uint32_t single_alu_op : 1;
342 					uint32_t reserved: 29;
343 					uint32_t process_ctx_flush: 1;
344 				};
345 				uint32_t u32all;
346 			} flags;
347 			uint32_t spi_gdbg_per_vmid_cntl;
348 			uint32_t tcp_watch_cntl[4];
349 			uint32_t trap_en;
350 		} set_shader_debugger;
351 
352 		struct {
353 			union {
354 				struct {
355 					uint32_t limit_single_process : 1;
356 					uint32_t enable_hws_logging_buffer : 1;
357 					uint32_t reserved : 30;
358 				};
359 				uint32_t all;
360 			} option;
361 			struct {
362 				uint32_t tdr_level;
363 				uint32_t tdr_delay;
364 			} tdr_config;
365 		} change_config;
366 	};
367 };
368 
369 struct amdgpu_mes_funcs {
370 	int (*add_hw_queue)(struct amdgpu_mes *mes,
371 			    struct mes_add_queue_input *input);
372 
373 	int (*remove_hw_queue)(struct amdgpu_mes *mes,
374 			       struct mes_remove_queue_input *input);
375 
376 	int (*map_legacy_queue)(struct amdgpu_mes *mes,
377 				struct mes_map_legacy_queue_input *input);
378 
379 	int (*unmap_legacy_queue)(struct amdgpu_mes *mes,
380 				  struct mes_unmap_legacy_queue_input *input);
381 
382 	int (*suspend_gang)(struct amdgpu_mes *mes,
383 			    struct mes_suspend_gang_input *input);
384 
385 	int (*resume_gang)(struct amdgpu_mes *mes,
386 			   struct mes_resume_gang_input *input);
387 
388 	int (*misc_op)(struct amdgpu_mes *mes,
389 		       struct mes_misc_op_input *input);
390 
391 	int (*reset_legacy_queue)(struct amdgpu_mes *mes,
392 				  struct mes_reset_legacy_queue_input *input);
393 
394 	int (*reset_hw_queue)(struct amdgpu_mes *mes,
395 			      struct mes_reset_queue_input *input);
396 };
397 
398 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
399 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev))
400 
401 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs);
402 
403 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe);
404 int amdgpu_mes_init(struct amdgpu_device *adev);
405 void amdgpu_mes_fini(struct amdgpu_device *adev);
406 
407 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
408 			      struct amdgpu_vm *vm);
409 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid);
410 
411 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
412 			struct amdgpu_mes_gang_properties *gprops,
413 			int *gang_id);
414 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id);
415 
416 int amdgpu_mes_suspend(struct amdgpu_device *adev);
417 int amdgpu_mes_resume(struct amdgpu_device *adev);
418 
419 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
420 			    struct amdgpu_mes_queue_properties *qprops,
421 			    int *queue_id);
422 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id);
423 int amdgpu_mes_reset_hw_queue(struct amdgpu_device *adev, int queue_id);
424 int amdgpu_mes_reset_hw_queue_mmio(struct amdgpu_device *adev, int queue_type,
425 				   int me_id, int pipe_id, int queue_id, int vmid);
426 
427 int amdgpu_mes_map_legacy_queue(struct amdgpu_device *adev,
428 				struct amdgpu_ring *ring);
429 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
430 				  struct amdgpu_ring *ring,
431 				  enum amdgpu_unmap_queues_action action,
432 				  u64 gpu_addr, u64 seq);
433 int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
434 				  struct amdgpu_ring *ring,
435 				  unsigned int vmid,
436 				  bool use_mmio);
437 
438 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
439 int amdgpu_mes_wreg(struct amdgpu_device *adev,
440 		    uint32_t reg, uint32_t val);
441 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
442 			uint32_t val, uint32_t mask);
443 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
444 				  uint32_t reg0, uint32_t reg1,
445 				  uint32_t ref, uint32_t mask);
446 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
447 				uint64_t process_context_addr,
448 				uint32_t spi_gdbg_per_vmid_cntl,
449 				const uint32_t *tcp_watch_cntl,
450 				uint32_t flags,
451 				bool trap_en);
452 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
453 				uint64_t process_context_addr);
454 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
455 			int queue_type, int idx,
456 			struct amdgpu_mes_ctx_data *ctx_data,
457 			struct amdgpu_ring **out);
458 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
459 			    struct amdgpu_ring *ring);
460 
461 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
462 						   enum amdgpu_mes_priority_level prio);
463 
464 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
465 				   struct amdgpu_mes_ctx_data *ctx_data);
466 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data);
467 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
468 				 struct amdgpu_vm *vm,
469 				 struct amdgpu_mes_ctx_data *ctx_data);
470 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
471 				   struct amdgpu_mes_ctx_data *ctx_data);
472 
473 int amdgpu_mes_self_test(struct amdgpu_device *adev);
474 
475 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev);
476 
477 /*
478  * MES lock can be taken in MMU notifiers.
479  *
480  * A bit more detail about why to set no-FS reclaim with MES lock:
481  *
482  * The purpose of the MMU notifier is to stop GPU access to memory so
483  * that the Linux VM subsystem can move pages around safely. This is
484  * done by preempting user mode queues for the affected process. When
485  * MES is used, MES lock needs to be taken to preempt the queues.
486  *
487  * The MMU notifier callback entry point in the driver is
488  * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from
489  * there is:
490  * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm ->
491  * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues
492  *
493  * The last part of the chain is a function pointer where we take the
494  * MES lock.
495  *
496  * The problem with taking locks in the MMU notifier is, that MMU
497  * notifiers can be called in reclaim-FS context. That's where the
498  * kernel frees up pages to make room for new page allocations under
499  * memory pressure. While we are running in reclaim-FS context, we must
500  * not trigger another memory reclaim operation because that would
501  * recursively reenter the reclaim code and cause a deadlock. The
502  * memalloc_nofs_save/restore calls guarantee that.
503  *
504  * In addition we also need to avoid lock dependencies on other locks taken
505  * under the MES lock, for example reservation locks. Here is a possible
506  * scenario of a deadlock:
507  * Thread A: takes and holds reservation lock | triggers reclaim-FS |
508  * MMU notifier | blocks trying to take MES lock
509  * Thread B: takes and holds MES lock | blocks trying to take reservation lock
510  *
511  * In this scenario Thread B gets involved in a deadlock even without
512  * triggering a reclaim-FS operation itself.
513  * To fix this and break the lock dependency chain you'd need to either:
514  * 1. protect reservation locks with memalloc_nofs_save/restore, or
515  * 2. avoid taking reservation locks under the MES lock.
516  *
517  * Reservation locks are taken all over the kernel in different subsystems, we
518  * have no control over them and their lock dependencies.So the only workable
519  * solution is to avoid taking other locks under the MES lock.
520  * As a result, make sure no reclaim-FS happens while holding this lock anywhere
521  * to prevent deadlocks when an MMU notifier runs in reclaim-FS context.
522  */
523 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes)
524 {
525 	mutex_lock(&mes->mutex_hidden);
526 	mes->saved_flags = memalloc_noreclaim_save();
527 }
528 
529 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes)
530 {
531 	memalloc_noreclaim_restore(mes->saved_flags);
532 	mutex_unlock(&mes->mutex_hidden);
533 }
534 
535 bool amdgpu_mes_suspend_resume_all_supported(struct amdgpu_device *adev);
536 
537 int amdgpu_mes_set_enforce_isolation(struct amdgpu_device *adev, uint32_t node_id, bool enable);
538 
539 #endif /* __AMDGPU_MES_H__ */
540