xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h (revision f2745dc0ba3dadd8fa2b2c33f48253d78e133a12)
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 enum amdgpu_mes_priority_level {
37 	AMDGPU_MES_PRIORITY_LEVEL_LOW       = 0,
38 	AMDGPU_MES_PRIORITY_LEVEL_NORMAL    = 1,
39 	AMDGPU_MES_PRIORITY_LEVEL_MEDIUM    = 2,
40 	AMDGPU_MES_PRIORITY_LEVEL_HIGH      = 3,
41 	AMDGPU_MES_PRIORITY_LEVEL_REALTIME  = 4,
42 	AMDGPU_MES_PRIORITY_NUM_LEVELS
43 };
44 
45 #define AMDGPU_MES_PROC_CTX_SIZE 0x1000 /* one page area */
46 #define AMDGPU_MES_GANG_CTX_SIZE 0x1000 /* one page area */
47 
48 struct amdgpu_mes_funcs;
49 
50 enum admgpu_mes_pipe {
51 	AMDGPU_MES_SCHED_PIPE = 0,
52 	AMDGPU_MES_KIQ_PIPE,
53 	AMDGPU_MAX_MES_PIPES = 2,
54 };
55 
56 struct amdgpu_mes {
57 	struct amdgpu_device            *adev;
58 
59 	struct mutex                    mutex_hidden;
60 
61 	struct idr                      pasid_idr;
62 	struct idr                      gang_id_idr;
63 	struct idr                      queue_id_idr;
64 	struct ida                      doorbell_ida;
65 
66 	spinlock_t                      queue_id_lock;
67 
68 	uint32_t                        total_max_queue;
69 	uint32_t                        doorbell_id_offset;
70 	uint32_t                        max_doorbell_slices;
71 
72 	uint64_t                        default_process_quantum;
73 	uint64_t                        default_gang_quantum;
74 
75 	struct amdgpu_ring              ring;
76 
77 	const struct firmware           *fw[AMDGPU_MAX_MES_PIPES];
78 
79 	/* mes ucode */
80 	struct amdgpu_bo		*ucode_fw_obj[AMDGPU_MAX_MES_PIPES];
81 	uint64_t			ucode_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
82 	uint32_t			*ucode_fw_ptr[AMDGPU_MAX_MES_PIPES];
83 	uint32_t                        ucode_fw_version[AMDGPU_MAX_MES_PIPES];
84 	uint64_t                        uc_start_addr[AMDGPU_MAX_MES_PIPES];
85 
86 	/* mes ucode data */
87 	struct amdgpu_bo		*data_fw_obj[AMDGPU_MAX_MES_PIPES];
88 	uint64_t			data_fw_gpu_addr[AMDGPU_MAX_MES_PIPES];
89 	uint32_t			*data_fw_ptr[AMDGPU_MAX_MES_PIPES];
90 	uint32_t                        data_fw_version[AMDGPU_MAX_MES_PIPES];
91 	uint64_t                        data_start_addr[AMDGPU_MAX_MES_PIPES];
92 
93 	/* eop gpu obj */
94 	struct amdgpu_bo		*eop_gpu_obj[AMDGPU_MAX_MES_PIPES];
95 	uint64_t                        eop_gpu_addr[AMDGPU_MAX_MES_PIPES];
96 
97 	void                            *mqd_backup[AMDGPU_MAX_MES_PIPES];
98 	struct amdgpu_irq_src	        irq[AMDGPU_MAX_MES_PIPES];
99 
100 	uint32_t                        vmid_mask_gfxhub;
101 	uint32_t                        vmid_mask_mmhub;
102 	uint32_t                        compute_hqd_mask[AMDGPU_MES_MAX_COMPUTE_PIPES];
103 	uint32_t                        gfx_hqd_mask[AMDGPU_MES_MAX_GFX_PIPES];
104 	uint32_t                        sdma_hqd_mask[AMDGPU_MES_MAX_SDMA_PIPES];
105 	uint32_t                        agreegated_doorbells[AMDGPU_MES_PRIORITY_NUM_LEVELS];
106 	uint32_t                        sch_ctx_offs;
107 	uint64_t			sch_ctx_gpu_addr;
108 	uint64_t			*sch_ctx_ptr;
109 	uint32_t			query_status_fence_offs;
110 	uint64_t			query_status_fence_gpu_addr;
111 	uint64_t			*query_status_fence_ptr;
112 	uint32_t			saved_flags;
113 
114 	/* initialize kiq pipe */
115 	int                             (*kiq_hw_init)(struct amdgpu_device *adev);
116 	int                             (*kiq_hw_fini)(struct amdgpu_device *adev);
117 
118 	/* ip specific functions */
119 	const struct amdgpu_mes_funcs   *funcs;
120 };
121 
122 struct amdgpu_mes_process {
123 	int			pasid;
124 	struct			amdgpu_vm *vm;
125 	uint64_t		pd_gpu_addr;
126 	struct amdgpu_bo 	*proc_ctx_bo;
127 	uint64_t 		proc_ctx_gpu_addr;
128 	void 			*proc_ctx_cpu_ptr;
129 	uint64_t 		process_quantum;
130 	struct 			list_head gang_list;
131 	uint32_t 		doorbell_index;
132 	unsigned long 		*doorbell_bitmap;
133 	struct mutex		doorbell_lock;
134 };
135 
136 struct amdgpu_mes_gang {
137 	int 				gang_id;
138 	int 				priority;
139 	int 				inprocess_gang_priority;
140 	int 				global_priority_level;
141 	struct list_head 		list;
142 	struct amdgpu_mes_process 	*process;
143 	struct amdgpu_bo 		*gang_ctx_bo;
144 	uint64_t 			gang_ctx_gpu_addr;
145 	void 				*gang_ctx_cpu_ptr;
146 	uint64_t 			gang_quantum;
147 	struct list_head 		queue_list;
148 };
149 
150 struct amdgpu_mes_queue {
151 	struct list_head 		list;
152 	struct amdgpu_mes_gang 		*gang;
153 	int 				queue_id;
154 	uint64_t 			doorbell_off;
155 	struct amdgpu_bo		*mqd_obj;
156 	void				*mqd_cpu_ptr;
157 	uint64_t 			mqd_gpu_addr;
158 	uint64_t 			wptr_gpu_addr;
159 	int 				queue_type;
160 	int 				paging;
161 	struct amdgpu_ring 		*ring;
162 };
163 
164 struct amdgpu_mes_queue_properties {
165 	int 			queue_type;
166 	uint64_t                hqd_base_gpu_addr;
167 	uint64_t                rptr_gpu_addr;
168 	uint64_t                wptr_gpu_addr;
169 	uint32_t                queue_size;
170 	uint64_t                eop_gpu_addr;
171 	uint32_t                hqd_pipe_priority;
172 	uint32_t                hqd_queue_priority;
173 	bool 			paging;
174 	struct amdgpu_ring 	*ring;
175 	/* out */
176 	uint64_t       		doorbell_off;
177 };
178 
179 struct amdgpu_mes_gang_properties {
180 	uint32_t 	priority;
181 	uint32_t 	gang_quantum;
182 	uint32_t 	inprocess_gang_priority;
183 	uint32_t 	priority_level;
184 	int 		global_priority_level;
185 };
186 
187 struct mes_add_queue_input {
188 	uint32_t	process_id;
189 	uint64_t	page_table_base_addr;
190 	uint64_t	process_va_start;
191 	uint64_t	process_va_end;
192 	uint64_t	process_quantum;
193 	uint64_t	process_context_addr;
194 	uint64_t	gang_quantum;
195 	uint64_t	gang_context_addr;
196 	uint32_t	inprocess_gang_priority;
197 	uint32_t	gang_global_priority_level;
198 	uint32_t	doorbell_offset;
199 	uint64_t	mqd_addr;
200 	uint64_t	wptr_addr;
201 	uint32_t	queue_type;
202 	uint32_t	paging;
203 	uint32_t        gws_base;
204 	uint32_t        gws_size;
205 	uint64_t	tba_addr;
206 	uint64_t	tma_addr;
207 };
208 
209 struct mes_remove_queue_input {
210 	uint32_t	doorbell_offset;
211 	uint64_t	gang_context_addr;
212 };
213 
214 struct mes_unmap_legacy_queue_input {
215 	enum amdgpu_unmap_queues_action    action;
216 	uint32_t                           queue_type;
217 	uint32_t                           doorbell_offset;
218 	uint32_t                           pipe_id;
219 	uint32_t                           queue_id;
220 	uint64_t                           trail_fence_addr;
221 	uint64_t                           trail_fence_data;
222 };
223 
224 struct mes_suspend_gang_input {
225 	bool		suspend_all_gangs;
226 	uint64_t	gang_context_addr;
227 	uint64_t	suspend_fence_addr;
228 	uint32_t	suspend_fence_value;
229 };
230 
231 struct mes_resume_gang_input {
232 	bool		resume_all_gangs;
233 	uint64_t	gang_context_addr;
234 };
235 
236 struct amdgpu_mes_funcs {
237 	int (*add_hw_queue)(struct amdgpu_mes *mes,
238 			    struct mes_add_queue_input *input);
239 
240 	int (*remove_hw_queue)(struct amdgpu_mes *mes,
241 			       struct mes_remove_queue_input *input);
242 
243 	int (*unmap_legacy_queue)(struct amdgpu_mes *mes,
244 				  struct mes_unmap_legacy_queue_input *input);
245 
246 	int (*suspend_gang)(struct amdgpu_mes *mes,
247 			    struct mes_suspend_gang_input *input);
248 
249 	int (*resume_gang)(struct amdgpu_mes *mes,
250 			   struct mes_resume_gang_input *input);
251 };
252 
253 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
254 #define amdgpu_mes_kiq_hw_fini(adev) (adev)->mes.kiq_hw_fini((adev))
255 
256 int amdgpu_mes_ctx_get_offs(struct amdgpu_ring *ring, unsigned int id_offs);
257 
258 int amdgpu_mes_init(struct amdgpu_device *adev);
259 void amdgpu_mes_fini(struct amdgpu_device *adev);
260 
261 int amdgpu_mes_create_process(struct amdgpu_device *adev, int pasid,
262 			      struct amdgpu_vm *vm);
263 void amdgpu_mes_destroy_process(struct amdgpu_device *adev, int pasid);
264 
265 int amdgpu_mes_add_gang(struct amdgpu_device *adev, int pasid,
266 			struct amdgpu_mes_gang_properties *gprops,
267 			int *gang_id);
268 int amdgpu_mes_remove_gang(struct amdgpu_device *adev, int gang_id);
269 
270 int amdgpu_mes_suspend(struct amdgpu_device *adev);
271 int amdgpu_mes_resume(struct amdgpu_device *adev);
272 
273 int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
274 			    struct amdgpu_mes_queue_properties *qprops,
275 			    int *queue_id);
276 int amdgpu_mes_remove_hw_queue(struct amdgpu_device *adev, int queue_id);
277 
278 int amdgpu_mes_unmap_legacy_queue(struct amdgpu_device *adev,
279 				  struct amdgpu_ring *ring,
280 				  enum amdgpu_unmap_queues_action action,
281 				  u64 gpu_addr, u64 seq);
282 
283 int amdgpu_mes_add_ring(struct amdgpu_device *adev, int gang_id,
284 			int queue_type, int idx,
285 			struct amdgpu_mes_ctx_data *ctx_data,
286 			struct amdgpu_ring **out);
287 void amdgpu_mes_remove_ring(struct amdgpu_device *adev,
288 			    struct amdgpu_ring *ring);
289 
290 int amdgpu_mes_ctx_alloc_meta_data(struct amdgpu_device *adev,
291 				   struct amdgpu_mes_ctx_data *ctx_data);
292 void amdgpu_mes_ctx_free_meta_data(struct amdgpu_mes_ctx_data *ctx_data);
293 int amdgpu_mes_ctx_map_meta_data(struct amdgpu_device *adev,
294 				 struct amdgpu_vm *vm,
295 				 struct amdgpu_mes_ctx_data *ctx_data);
296 
297 int amdgpu_mes_self_test(struct amdgpu_device *adev);
298 
299 int amdgpu_mes_alloc_process_doorbells(struct amdgpu_device *adev,
300 					unsigned int *doorbell_index);
301 void amdgpu_mes_free_process_doorbells(struct amdgpu_device *adev,
302 					unsigned int doorbell_index);
303 unsigned int amdgpu_mes_get_doorbell_dw_offset_in_bar(
304 					struct amdgpu_device *adev,
305 					uint32_t doorbell_index,
306 					unsigned int doorbell_id);
307 int amdgpu_mes_doorbell_process_slice(struct amdgpu_device *adev);
308 
309 /*
310  * MES lock can be taken in MMU notifiers.
311  *
312  * A bit more detail about why to set no-FS reclaim with MES lock:
313  *
314  * The purpose of the MMU notifier is to stop GPU access to memory so
315  * that the Linux VM subsystem can move pages around safely. This is
316  * done by preempting user mode queues for the affected process. When
317  * MES is used, MES lock needs to be taken to preempt the queues.
318  *
319  * The MMU notifier callback entry point in the driver is
320  * amdgpu_mn_invalidate_range_start_hsa. The relevant call chain from
321  * there is:
322  * amdgpu_amdkfd_evict_userptr -> kgd2kfd_quiesce_mm ->
323  * kfd_process_evict_queues -> pdd->dev->dqm->ops.evict_process_queues
324  *
325  * The last part of the chain is a function pointer where we take the
326  * MES lock.
327  *
328  * The problem with taking locks in the MMU notifier is, that MMU
329  * notifiers can be called in reclaim-FS context. That's where the
330  * kernel frees up pages to make room for new page allocations under
331  * memory pressure. While we are running in reclaim-FS context, we must
332  * not trigger another memory reclaim operation because that would
333  * recursively reenter the reclaim code and cause a deadlock. The
334  * memalloc_nofs_save/restore calls guarantee that.
335  *
336  * In addition we also need to avoid lock dependencies on other locks taken
337  * under the MES lock, for example reservation locks. Here is a possible
338  * scenario of a deadlock:
339  * Thread A: takes and holds reservation lock | triggers reclaim-FS |
340  * MMU notifier | blocks trying to take MES lock
341  * Thread B: takes and holds MES lock | blocks trying to take reservation lock
342  *
343  * In this scenario Thread B gets involved in a deadlock even without
344  * triggering a reclaim-FS operation itself.
345  * To fix this and break the lock dependency chain you'd need to either:
346  * 1. protect reservation locks with memalloc_nofs_save/restore, or
347  * 2. avoid taking reservation locks under the MES lock.
348  *
349  * Reservation locks are taken all over the kernel in different subsystems, we
350  * have no control over them and their lock dependencies.So the only workable
351  * solution is to avoid taking other locks under the MES lock.
352  * As a result, make sure no reclaim-FS happens while holding this lock anywhere
353  * to prevent deadlocks when an MMU notifier runs in reclaim-FS context.
354  */
355 static inline void amdgpu_mes_lock(struct amdgpu_mes *mes)
356 {
357 	mutex_lock(&mes->mutex_hidden);
358 	mes->saved_flags = memalloc_noreclaim_save();
359 }
360 
361 static inline void amdgpu_mes_unlock(struct amdgpu_mes *mes)
362 {
363 	memalloc_noreclaim_restore(mes->saved_flags);
364 	mutex_unlock(&mes->mutex_hidden);
365 }
366 #endif /* __AMDGPU_MES_H__ */
367