xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h (revision ef9226cd56b718c79184a3466d32984a51cb449c)
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 
44 enum amdgpu_mes_priority_level {
45 	AMDGPU_MES_PRIORITY_LEVEL_LOW       = 0,
46 	AMDGPU_MES_PRIORITY_LEVEL_NORMAL    = 1,
47 	AMDGPU_MES_PRIORITY_LEVEL_MEDIUM    = 2,
48 	AMDGPU_MES_PRIORITY_LEVEL_HIGH      = 3,
49 	AMDGPU_MES_PRIORITY_LEVEL_REALTIME  = 4,
50 	AMDGPU_MES_PRIORITY_NUM_LEVELS
51 };
52 
53 #define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */
54 #define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */
55 #define AMDGPU_MES_LOG_BUFFER_SIZE 0x4000 /* Maximu log buffer size for MES */
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 
80 	uint32_t                        total_max_queue;
81 	uint32_t                        max_doorbell_slices;
82 
83 	uint64_t                        default_process_quantum;
84 	uint64_t                        default_gang_quantum;
85 
86 	struct amdgpu_ring              ring;
87 	spinlock_t                      ring_lock;
88 
89 	const struct firmware           *fw[AMDGPU_MAX_MES_PIPES];
90 
91 	/* mes ucode */
92 	struct amdgpu_bo		*ucode_fw_obj[AMDGPU_MAX_MES_PIPES];
93 	uint64_t			ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
94 	uint32_t			*ucode_fw_ptr[AMDGPU_MAX_MES_PIPES];
95 	uint64_t                        uc_start_addr[AMDGPU_MAX_MES_PIPES];
96 
97 	/* mes ucode data */
98 	struct amdgpu_bo		*data_fw_obj[AMDGPU_MAX_MES_PIPES];
99 	uint64_t			data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
100 	uint32_t			*data_fw_ptr[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                        aggregated_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 	/* MES doorbells */
133 	uint32_t			db_start_dw_offset;
134 	uint32_t			num_mes_dbs;
135 	unsigned long			*doorbell_bitmap;
136 
137 	/* MES event log buffer */
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 
146 struct amdgpu_mes_process {
147 	int			pasid;
148 	struct			amdgpu_vm *vm;
149 	uint64_t		pd_gpu_addr;
150 	struct amdgpu_bo 	*proc_ctx_bo;
151 	uint64_t 		proc_ctx_gpu_addr;
152 	void 			*proc_ctx_cpu_ptr;
153 	uint64_t 		process_quantum;
154 	struct 			list_head gang_list;
155 	uint32_t 		doorbell_index;
156 	struct mutex		doorbell_lock;
157 };
158 
159 struct amdgpu_mes_gang {
160 	int 				gang_id;
161 	int 				priority;
162 	int 				inprocess_gang_priority;
163 	int 				global_priority_level;
164 	struct list_head 		list;
165 	struct amdgpu_mes_process 	*process;
166 	struct amdgpu_bo 		*gang_ctx_bo;
167 	uint64_t 			gang_ctx_gpu_addr;
168 	void 				*gang_ctx_cpu_ptr;
169 	uint64_t 			gang_quantum;
170 	struct list_head 		queue_list;
171 };
172 
173 struct amdgpu_mes_queue {
174 	struct list_head 		list;
175 	struct amdgpu_mes_gang 		*gang;
176 	int 				queue_id;
177 	uint64_t 			doorbell_off;
178 	struct amdgpu_bo		*mqd_obj;
179 	void				*mqd_cpu_ptr;
180 	uint64_t 			mqd_gpu_addr;
181 	uint64_t 			wptr_gpu_addr;
182 	int 				queue_type;
183 	int 				paging;
184 	struct amdgpu_ring 		*ring;
185 };
186 
187 struct amdgpu_mes_queue_properties {
188 	int 			queue_type;
189 	uint64_t                hqd_base_gpu_addr;
190 	uint64_t                rptr_gpu_addr;
191 	uint64_t                wptr_gpu_addr;
192 	uint64_t                wptr_mc_addr;
193 	uint32_t                queue_size;
194 	uint64_t                eop_gpu_addr;
195 	uint32_t                hqd_pipe_priority;
196 	uint32_t                hqd_queue_priority;
197 	bool 			paging;
198 	struct amdgpu_ring 	*ring;
199 	/* out */
200 	uint64_t       		doorbell_off;
201 };
202 
203 struct amdgpu_mes_gang_properties {
204 	uint32_t 	priority;
205 	uint32_t 	gang_quantum;
206 	uint32_t 	inprocess_gang_priority;
207 	uint32_t 	priority_level;
208 	int 		global_priority_level;
209 };
210 
211 struct mes_add_queue_input {
212 	uint32_t	process_id;
213 	uint64_t	page_table_base_addr;
214 	uint64_t	process_va_start;
215 	uint64_t	process_va_end;
216 	uint64_t	process_quantum;
217 	uint64_t	process_context_addr;
218 	uint64_t	gang_quantum;
219 	uint64_t	gang_context_addr;
220 	uint32_t	inprocess_gang_priority;
221 	uint32_t	gang_global_priority_level;
222 	uint32_t	doorbell_offset;
223 	uint64_t	mqd_addr;
224 	uint64_t	wptr_addr;
225 	uint64_t	wptr_mc_addr;
226 	uint32_t	queue_type;
227 	uint32_t	paging;
228 	uint32_t        gws_base;
229 	uint32_t        gws_size;
230 	uint64_t	tba_addr;
231 	uint64_t	tma_addr;
232 	uint32_t	trap_en;
233 	uint32_t	skip_process_ctx_clear;
234 	uint32_t	is_kfd_process;
235 	uint32_t	is_aql_queue;
236 	uint32_t	queue_size;
237 	uint32_t	exclusively_scheduled;
238 };
239 
240 struct mes_remove_queue_input {
241 	uint32_t	doorbell_offset;
242 	uint64_t	gang_context_addr;
243 };
244 
245 struct mes_unmap_legacy_queue_input {
246 	enum amdgpu_unmap_queues_action    action;
247 	uint32_t                           queue_type;
248 	uint32_t                           doorbell_offset;
249 	uint32_t                           pipe_id;
250 	uint32_t                           queue_id;
251 	uint64_t                           trail_fence_addr;
252 	uint64_t                           trail_fence_data;
253 };
254 
255 struct mes_suspend_gang_input {
256 	bool		suspend_all_gangs;
257 	uint64_t	gang_context_addr;
258 	uint64_t	suspend_fence_addr;
259 	uint32_t	suspend_fence_value;
260 };
261 
262 struct mes_resume_gang_input {
263 	bool		resume_all_gangs;
264 	uint64_t	gang_context_addr;
265 };
266 
267 enum mes_misc_opcode {
268 	MES_MISC_OP_WRITE_REG,
269 	MES_MISC_OP_READ_REG,
270 	MES_MISC_OP_WRM_REG_WAIT,
271 	MES_MISC_OP_WRM_REG_WR_WAIT,
272 	MES_MISC_OP_SET_SHADER_DEBUGGER,
273 };
274 
275 struct mes_misc_op_input {
276 	enum mes_misc_opcode op;
277 
278 	union {
279 		struct {
280 			uint32_t                  reg_offset;
281 			uint64_t                  buffer_addr;
282 		} read_reg;
283 
284 		struct {
285 			uint32_t                  reg_offset;
286 			uint32_t                  reg_value;
287 		} write_reg;
288 
289 		struct {
290 			uint32_t                   ref;
291 			uint32_t                   mask;
292 			uint32_t                   reg0;
293 			uint32_t                   reg1;
294 		} wrm_reg;
295 
296 		struct {
297 			uint64_t process_context_addr;
298 			union {
299 				struct {
300 					uint32_t single_memop : 1;
301 					uint32_t single_alu_op : 1;
302 					uint32_t reserved: 29;
303 					uint32_t process_ctx_flush: 1;
304 				};
305 				uint32_t u32all;
306 			} flags;
307 			uint32_t spi_gdbg_per_vmid_cntl;
308 			uint32_t tcp_watch_cntl[4];
309 			uint32_t trap_en;
310 		} set_shader_debugger;
311 	};
312 };
313 
314 struct amdgpu_mes_funcs {
315 	int (*add_hw_queue)(struct amdgpu_mes *mes,
316 			    struct mes_add_queue_input *input);
317 
318 	int (*remove_hw_queue)(struct amdgpu_mes *mes,
319 			       struct mes_remove_queue_input *input);
320 
321 	int (*unmap_legacy_queue)(struct amdgpu_mes *mes,
322 				  struct mes_unmap_legacy_queue_input *input);
323 
324 	int (*suspend_gang)(struct amdgpu_mes *mes,
325 			    struct mes_suspend_gang_input *input);
326 
327 	int (*resume_gang)(struct amdgpu_mes *mes,
328 			   struct mes_resume_gang_input *input);
329 
330 	int (*misc_op)(struct amdgpu_mes *mes,
331 		       struct mes_misc_op_input *input);
332 };
333 
334 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
335 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev))
336 
337 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs);
338 
339 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe);
340 int amdgpu_mes_init(struct amdgpu_device *adev);
341 void amdgpu_mes_fini(struct amdgpu_device *adev);
342 
343 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
344 			      struct amdgpu_vm *vm);
345 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid);
346 
347 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
348 			struct amdgpu_mes_gang_properties *gprops,
349 			int *gang_id);
350 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id);
351 
352 int amdgpu_mes_suspend(struct amdgpu_device *adev);
353 int amdgpu_mes_resume(struct amdgpu_device *adev);
354 
355 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
356 			    struct amdgpu_mes_queue_properties *qprops,
357 			    int *queue_id);
358 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id);
359 
360 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
361 				  struct amdgpu_ring *ring,
362 				  enum amdgpu_unmap_queues_action action,
363 				  u64 gpu_addr, u64 seq);
364 
365 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
366 int amdgpu_mes_wreg(struct amdgpu_device *adev,
367 		    uint32_t reg, uint32_t val);
368 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
369 			uint32_t val, uint32_t mask);
370 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
371 				  uint32_t reg0, uint32_t reg1,
372 				  uint32_t ref, uint32_t mask);
373 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
374 				uint64_t process_context_addr,
375 				uint32_t spi_gdbg_per_vmid_cntl,
376 				const uint32_t *tcp_watch_cntl,
377 				uint32_t flags,
378 				bool trap_en);
379 int amdgpu_mes_flush_shader_debugger(struct amdgpu_device *adev,
380 				uint64_t process_context_addr);
381 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
382 			int queue_type, int idx,
383 			struct amdgpu_mes_ctx_data *ctx_data,
384 			struct amdgpu_ring **out);
385 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
386 			    struct amdgpu_ring *ring);
387 
388 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
389 						   enum amdgpu_mes_priority_level prio);
390 
391 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
392 				   struct amdgpu_mes_ctx_data *ctx_data);
393 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data);
394 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
395 				 struct amdgpu_vm *vm,
396 				 struct amdgpu_mes_ctx_data *ctx_data);
397 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
398 				   struct amdgpu_mes_ctx_data *ctx_data);
399 
400 int amdgpu_mes_self_test(struct amdgpu_device *adev);
401 
402 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev);
403 
404 /*
405  * MES lock can be taken in MMU notifiers.
406  *
407  * A bit more detail about why to set no-FS reclaim with MES lock:
408  *
409  * The purpose of the MMU notifier is to stop GPU access to memory so
410  * that the Linux VM subsystem can move pages around safely. This is
411  * done by preempting user mode queues for the affected process. When
412  * MES is used, MES lock needs to be taken to preempt the queues.
413  *
414  * The MMU notifier callback entry point in the driver is
415  * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from
416  * there is:
417  * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm ->
418  * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues
419  *
420  * The last part of the chain is a function pointer where we take the
421  * MES lock.
422  *
423  * The problem with taking locks in the MMU notifier is, that MMU
424  * notifiers can be called in reclaim-FS context. That's where the
425  * kernel frees up pages to make room for new page allocations under
426  * memory pressure. While we are running in reclaim-FS context, we must
427  * not trigger another memory reclaim operation because that would
428  * recursively reenter the reclaim code and cause a deadlock. The
429  * memalloc_nofs_save/restore calls guarantee that.
430  *
431  * In addition we also need to avoid lock dependencies on other locks taken
432  * under the MES lock, for example reservation locks. Here is a possible
433  * scenario of a deadlock:
434  * Thread A: takes and holds reservation lock | triggers reclaim-FS |
435  * MMU notifier | blocks trying to take MES lock
436  * Thread B: takes and holds MES lock | blocks trying to take reservation lock
437  *
438  * In this scenario Thread B gets involved in a deadlock even without
439  * triggering a reclaim-FS operation itself.
440  * To fix this and break the lock dependency chain you'd need to either:
441  * 1. protect reservation locks with memalloc_nofs_save/restore, or
442  * 2. avoid taking reservation locks under the MES lock.
443  *
444  * Reservation locks are taken all over the kernel in different subsystems, we
445  * have no control over them and their lock dependencies.So the only workable
446  * solution is to avoid taking other locks under the MES lock.
447  * As a result, make sure no reclaim-FS happens while holding this lock anywhere
448  * to prevent deadlocks when an MMU notifier runs in reclaim-FS context.
449  */
450 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes)
451 {
452 	mutex_lock(&mes->mutex_hidden);
453 	mes->saved_flags = memalloc_noreclaim_save();
454 }
455 
456 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes)
457 {
458 	memalloc_noreclaim_restore(mes->saved_flags);
459 	mutex_unlock(&mes->mutex_hidden);
460 }
461 #endif /* __AMDGPU_MES_H__ */
462