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 
56 struct amdgpu_mes_funcs;
57 
58 enum admgpu_mes_pipe {
59 	AMDGPU_MES_SCHED_PIPE = 0,
60 	AMDGPU_MES_KIQ_PIPE,
61 	AMDGPU_MAX_MES_PIPES = 2,
62 };
63 
64 struct amdgpu_mes {
65 	struct amdgpu_device            *adev;
66 
67 	struct mutex                    mutex_hidden;
68 
69 	struct idr                      pasid_idr;
70 	struct idr                      gang_id_idr;
71 	struct idr                      queue_id_idr;
72 	struct ida                      doorbell_ida;
73 
74 	spinlock_t                      queue_id_lock;
75 
76 	uint32_t			sched_version;
77 	uint32_t			kiq_version;
78 
79 	uint32_t                        total_max_queue;
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 	spinlock_t                      ring_lock;
87 
88 	const struct firmware           *fw[AMDGPU_MAX_MES_PIPES];
89 
90 	/* mes ucode */
91 	struct amdgpu_bo		*ucode_fw_obj[AMDGPU_MAX_MES_PIPES];
92 	uint64_t			ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
93 	uint32_t			*ucode_fw_ptr[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 	uint64_t                        data_start_addr[AMDGPU_MAX_MES_PIPES];
101 
102 	/* eop gpu obj */
103 	struct amdgpu_bo		*eop_gpu_obj[AMDGPU_MAX_MES_PIPES];
104 	uint64_t                        eop_gpu_addr[AMDGPU_MAX_MES_PIPES];
105 
106 	void                            *mqd_backup[AMDGPU_MAX_MES_PIPES];
107 	struct amdgpu_irq_src	        irq[AMDGPU_MAX_MES_PIPES];
108 
109 	uint32_t                        vmid_mask_gfxhub;
110 	uint32_t                        vmid_mask_mmhub;
111 	uint32_t                        compute_hqd_mask[AMDGPU_MES_MAX_COMPUTE_PIPES];
112 	uint32_t                        gfx_hqd_mask[AMDGPU_MES_MAX_GFX_PIPES];
113 	uint32_t                        sdma_hqd_mask[AMDGPU_MES_MAX_SDMA_PIPES];
114 	uint32_t                        aggregated_doorbells[AMDGPU_MES_PRIORITY_NUM_LEVELS];
115 	uint32_t                        sch_ctx_offs;
116 	uint64_t			sch_ctx_gpu_addr;
117 	uint64_t			*sch_ctx_ptr;
118 	uint32_t			query_status_fence_offs;
119 	uint64_t			query_status_fence_gpu_addr;
120 	uint64_t			*query_status_fence_ptr;
121 	uint32_t                        read_val_offs;
122 	uint64_t			read_val_gpu_addr;
123 	uint32_t			*read_val_ptr;
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 	/* ip specific functions */
137 	const struct amdgpu_mes_funcs   *funcs;
138 };
139 
140 struct amdgpu_mes_process {
141 	int			pasid;
142 	struct			amdgpu_vm *vm;
143 	uint64_t		pd_gpu_addr;
144 	struct amdgpu_bo 	*proc_ctx_bo;
145 	uint64_t 		proc_ctx_gpu_addr;
146 	void 			*proc_ctx_cpu_ptr;
147 	uint64_t 		process_quantum;
148 	struct 			list_head gang_list;
149 	uint32_t 		doorbell_index;
150 	struct mutex		doorbell_lock;
151 };
152 
153 struct amdgpu_mes_gang {
154 	int 				gang_id;
155 	int 				priority;
156 	int 				inprocess_gang_priority;
157 	int 				global_priority_level;
158 	struct list_head 		list;
159 	struct amdgpu_mes_process 	*process;
160 	struct amdgpu_bo 		*gang_ctx_bo;
161 	uint64_t 			gang_ctx_gpu_addr;
162 	void 				*gang_ctx_cpu_ptr;
163 	uint64_t 			gang_quantum;
164 	struct list_head 		queue_list;
165 };
166 
167 struct amdgpu_mes_queue {
168 	struct list_head 		list;
169 	struct amdgpu_mes_gang 		*gang;
170 	int 				queue_id;
171 	uint64_t 			doorbell_off;
172 	struct amdgpu_bo		*mqd_obj;
173 	void				*mqd_cpu_ptr;
174 	uint64_t 			mqd_gpu_addr;
175 	uint64_t 			wptr_gpu_addr;
176 	int 				queue_type;
177 	int 				paging;
178 	struct amdgpu_ring 		*ring;
179 };
180 
181 struct amdgpu_mes_queue_properties {
182 	int 			queue_type;
183 	uint64_t                hqd_base_gpu_addr;
184 	uint64_t                rptr_gpu_addr;
185 	uint64_t                wptr_gpu_addr;
186 	uint64_t                wptr_mc_addr;
187 	uint32_t                queue_size;
188 	uint64_t                eop_gpu_addr;
189 	uint32_t                hqd_pipe_priority;
190 	uint32_t                hqd_queue_priority;
191 	bool 			paging;
192 	struct amdgpu_ring 	*ring;
193 	/* out */
194 	uint64_t       		doorbell_off;
195 };
196 
197 struct amdgpu_mes_gang_properties {
198 	uint32_t 	priority;
199 	uint32_t 	gang_quantum;
200 	uint32_t 	inprocess_gang_priority;
201 	uint32_t 	priority_level;
202 	int 		global_priority_level;
203 };
204 
205 struct mes_add_queue_input {
206 	uint32_t	process_id;
207 	uint64_t	page_table_base_addr;
208 	uint64_t	process_va_start;
209 	uint64_t	process_va_end;
210 	uint64_t	process_quantum;
211 	uint64_t	process_context_addr;
212 	uint64_t	gang_quantum;
213 	uint64_t	gang_context_addr;
214 	uint32_t	inprocess_gang_priority;
215 	uint32_t	gang_global_priority_level;
216 	uint32_t	doorbell_offset;
217 	uint64_t	mqd_addr;
218 	uint64_t	wptr_addr;
219 	uint64_t	wptr_mc_addr;
220 	uint32_t	queue_type;
221 	uint32_t	paging;
222 	uint32_t        gws_base;
223 	uint32_t        gws_size;
224 	uint64_t	tba_addr;
225 	uint64_t	tma_addr;
226 	uint32_t	trap_en;
227 	uint32_t	skip_process_ctx_clear;
228 	uint32_t	is_kfd_process;
229 	uint32_t	is_aql_queue;
230 	uint32_t	queue_size;
231 	uint32_t	exclusively_scheduled;
232 };
233 
234 struct mes_remove_queue_input {
235 	uint32_t	doorbell_offset;
236 	uint64_t	gang_context_addr;
237 };
238 
239 struct mes_unmap_legacy_queue_input {
240 	enum amdgpu_unmap_queues_action    action;
241 	uint32_t                           queue_type;
242 	uint32_t                           doorbell_offset;
243 	uint32_t                           pipe_id;
244 	uint32_t                           queue_id;
245 	uint64_t                           trail_fence_addr;
246 	uint64_t                           trail_fence_data;
247 };
248 
249 struct mes_suspend_gang_input {
250 	bool		suspend_all_gangs;
251 	uint64_t	gang_context_addr;
252 	uint64_t	suspend_fence_addr;
253 	uint32_t	suspend_fence_value;
254 };
255 
256 struct mes_resume_gang_input {
257 	bool		resume_all_gangs;
258 	uint64_t	gang_context_addr;
259 };
260 
261 enum mes_misc_opcode {
262 	MES_MISC_OP_WRITE_REG,
263 	MES_MISC_OP_READ_REG,
264 	MES_MISC_OP_WRM_REG_WAIT,
265 	MES_MISC_OP_WRM_REG_WR_WAIT,
266 	MES_MISC_OP_SET_SHADER_DEBUGGER,
267 };
268 
269 struct mes_misc_op_input {
270 	enum mes_misc_opcode op;
271 
272 	union {
273 		struct {
274 			uint32_t                  reg_offset;
275 			uint64_t                  buffer_addr;
276 		} read_reg;
277 
278 		struct {
279 			uint32_t                  reg_offset;
280 			uint32_t                  reg_value;
281 		} write_reg;
282 
283 		struct {
284 			uint32_t                   ref;
285 			uint32_t                   mask;
286 			uint32_t                   reg0;
287 			uint32_t                   reg1;
288 		} wrm_reg;
289 
290 		struct {
291 			uint64_t process_context_addr;
292 			union {
293 				struct {
294 					uint64_t single_memop : 1;
295 					uint64_t single_alu_op : 1;
296 					uint64_t reserved: 30;
297 				};
298 				uint32_t u32all;
299 			} flags;
300 			uint32_t spi_gdbg_per_vmid_cntl;
301 			uint32_t tcp_watch_cntl[4];
302 			uint32_t trap_en;
303 		} set_shader_debugger;
304 	};
305 };
306 
307 struct amdgpu_mes_funcs {
308 	int (*add_hw_queue)(struct amdgpu_mes *mes,
309 			    struct mes_add_queue_input *input);
310 
311 	int (*remove_hw_queue)(struct amdgpu_mes *mes,
312 			       struct mes_remove_queue_input *input);
313 
314 	int (*unmap_legacy_queue)(struct amdgpu_mes *mes,
315 				  struct mes_unmap_legacy_queue_input *input);
316 
317 	int (*suspend_gang)(struct amdgpu_mes *mes,
318 			    struct mes_suspend_gang_input *input);
319 
320 	int (*resume_gang)(struct amdgpu_mes *mes,
321 			   struct mes_resume_gang_input *input);
322 
323 	int (*misc_op)(struct amdgpu_mes *mes,
324 		       struct mes_misc_op_input *input);
325 };
326 
327 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
328 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev))
329 
330 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs);
331 
332 int amdgpu_mes_init_microcode(struct amdgpu_device *adev, int pipe);
333 int amdgpu_mes_init(struct amdgpu_device *adev);
334 void amdgpu_mes_fini(struct amdgpu_device *adev);
335 
336 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
337 			      struct amdgpu_vm *vm);
338 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid);
339 
340 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
341 			struct amdgpu_mes_gang_properties *gprops,
342 			int *gang_id);
343 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id);
344 
345 int amdgpu_mes_suspend(struct amdgpu_device *adev);
346 int amdgpu_mes_resume(struct amdgpu_device *adev);
347 
348 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
349 			    struct amdgpu_mes_queue_properties *qprops,
350 			    int *queue_id);
351 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id);
352 
353 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
354 				  struct amdgpu_ring *ring,
355 				  enum amdgpu_unmap_queues_action action,
356 				  u64 gpu_addr, u64 seq);
357 
358 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
359 int amdgpu_mes_wreg(struct amdgpu_device *adev,
360 		    uint32_t reg, uint32_t val);
361 int amdgpu_mes_reg_wait(struct amdgpu_device *adev, uint32_t reg,
362 			uint32_t val, uint32_t mask);
363 int amdgpu_mes_reg_write_reg_wait(struct amdgpu_device *adev,
364 				  uint32_t reg0, uint32_t reg1,
365 				  uint32_t ref, uint32_t mask);
366 int amdgpu_mes_set_shader_debugger(struct amdgpu_device *adev,
367 				uint64_t process_context_addr,
368 				uint32_t spi_gdbg_per_vmid_cntl,
369 				const uint32_t *tcp_watch_cntl,
370 				uint32_t flags,
371 				bool trap_en);
372 
373 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
374 			int queue_type, int idx,
375 			struct amdgpu_mes_ctx_data *ctx_data,
376 			struct amdgpu_ring **out);
377 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
378 			    struct amdgpu_ring *ring);
379 
380 uint32_t amdgpu_mes_get_aggregated_doorbell_index(struct amdgpu_device *adev,
381 						   enum amdgpu_mes_priority_level prio);
382 
383 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
384 				   struct amdgpu_mes_ctx_data *ctx_data);
385 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data);
386 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
387 				 struct amdgpu_vm *vm,
388 				 struct amdgpu_mes_ctx_data *ctx_data);
389 int amdgpu_mes_ctx_unmap_meta_data(struct amdgpu_device *adev,
390 				   struct amdgpu_mes_ctx_data *ctx_data);
391 
392 int amdgpu_mes_self_test(struct amdgpu_device *adev);
393 
394 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev);
395 
396 /*
397  * MES lock can be taken in MMU notifiers.
398  *
399  * A bit more detail about why to set no-FS reclaim with MES lock:
400  *
401  * The purpose of the MMU notifier is to stop GPU access to memory so
402  * that the Linux VM subsystem can move pages around safely. This is
403  * done by preempting user mode queues for the affected process. When
404  * MES is used, MES lock needs to be taken to preempt the queues.
405  *
406  * The MMU notifier callback entry point in the driver is
407  * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from
408  * there is:
409  * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm ->
410  * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues
411  *
412  * The last part of the chain is a function pointer where we take the
413  * MES lock.
414  *
415  * The problem with taking locks in the MMU notifier is, that MMU
416  * notifiers can be called in reclaim-FS context. That's where the
417  * kernel frees up pages to make room for new page allocations under
418  * memory pressure. While we are running in reclaim-FS context, we must
419  * not trigger another memory reclaim operation because that would
420  * recursively reenter the reclaim code and cause a deadlock. The
421  * memalloc_nofs_save/restore calls guarantee that.
422  *
423  * In addition we also need to avoid lock dependencies on other locks taken
424  * under the MES lock, for example reservation locks. Here is a possible
425  * scenario of a deadlock:
426  * Thread A: takes and holds reservation lock | triggers reclaim-FS |
427  * MMU notifier | blocks trying to take MES lock
428  * Thread B: takes and holds MES lock | blocks trying to take reservation lock
429  *
430  * In this scenario Thread B gets involved in a deadlock even without
431  * triggering a reclaim-FS operation itself.
432  * To fix this and break the lock dependency chain you'd need to either:
433  * 1. protect reservation locks with memalloc_nofs_save/restore, or
434  * 2. avoid taking reservation locks under the MES lock.
435  *
436  * Reservation locks are taken all over the kernel in different subsystems, we
437  * have no control over them and their lock dependencies.So the only workable
438  * solution is to avoid taking other locks under the MES lock.
439  * As a result, make sure no reclaim-FS happens while holding this lock anywhere
440  * to prevent deadlocks when an MMU notifier runs in reclaim-FS context.
441  */
442 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes)
443 {
444 	mutex_lock(&mes->mutex_hidden);
445 	mes->saved_flags = memalloc_noreclaim_save();
446 }
447 
448 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes)
449 {
450 	memalloc_noreclaim_restore(mes->saved_flags);
451 	mutex_unlock(&mes->mutex_hidden);
452 }
453 #endif /* __AMDGPU_MES_H__ */
454