1 /*
2  * Copyright 2016 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  * Authors: Christian König
23  */
24 #ifndef __AMDGPU_RING_H__
25 #define __AMDGPU_RING_H__
26 
27 #include <drm/amdgpu_drm.h>
28 #include <drm/gpu_scheduler.h>
29 #include <drm/drm_print.h>
30 
31 struct amdgpu_device;
32 struct amdgpu_ring;
33 struct amdgpu_ib;
34 struct amdgpu_cs_parser;
35 struct amdgpu_job;
36 struct amdgpu_vm;
37 
38 /* max number of rings */
39 #define AMDGPU_MAX_RINGS		28
40 #define AMDGPU_MAX_HWIP_RINGS		8
41 #define AMDGPU_MAX_GFX_RINGS		2
42 #define AMDGPU_MAX_COMPUTE_RINGS	8
43 #define AMDGPU_MAX_VCE_RINGS		3
44 #define AMDGPU_MAX_UVD_ENC_RINGS	2
45 
46 enum amdgpu_ring_priority_level {
47 	AMDGPU_RING_PRIO_0,
48 	AMDGPU_RING_PRIO_1,
49 	AMDGPU_RING_PRIO_DEFAULT = 1,
50 	AMDGPU_RING_PRIO_2,
51 	AMDGPU_RING_PRIO_MAX
52 };
53 
54 /* some special values for the owner field */
55 #define AMDGPU_FENCE_OWNER_UNDEFINED	((void *)0ul)
56 #define AMDGPU_FENCE_OWNER_VM		((void *)1ul)
57 #define AMDGPU_FENCE_OWNER_KFD		((void *)2ul)
58 
59 #define AMDGPU_FENCE_FLAG_64BIT         (1 << 0)
60 #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
61 #define AMDGPU_FENCE_FLAG_TC_WB_ONLY    (1 << 2)
62 
63 #define to_amdgpu_ring(s) container_of((s), struct amdgpu_ring, sched)
64 
65 #define AMDGPU_IB_POOL_SIZE	(1024 * 1024)
66 
67 enum amdgpu_ring_type {
68 	AMDGPU_RING_TYPE_GFX		= AMDGPU_HW_IP_GFX,
69 	AMDGPU_RING_TYPE_COMPUTE	= AMDGPU_HW_IP_COMPUTE,
70 	AMDGPU_RING_TYPE_SDMA		= AMDGPU_HW_IP_DMA,
71 	AMDGPU_RING_TYPE_UVD		= AMDGPU_HW_IP_UVD,
72 	AMDGPU_RING_TYPE_VCE		= AMDGPU_HW_IP_VCE,
73 	AMDGPU_RING_TYPE_UVD_ENC	= AMDGPU_HW_IP_UVD_ENC,
74 	AMDGPU_RING_TYPE_VCN_DEC	= AMDGPU_HW_IP_VCN_DEC,
75 	AMDGPU_RING_TYPE_VCN_ENC	= AMDGPU_HW_IP_VCN_ENC,
76 	AMDGPU_RING_TYPE_VCN_JPEG	= AMDGPU_HW_IP_VCN_JPEG,
77 	AMDGPU_RING_TYPE_KIQ,
78 	AMDGPU_RING_TYPE_MES
79 };
80 
81 enum amdgpu_ib_pool_type {
82 	/* Normal submissions to the top of the pipeline. */
83 	AMDGPU_IB_POOL_DELAYED,
84 	/* Immediate submissions to the bottom of the pipeline. */
85 	AMDGPU_IB_POOL_IMMEDIATE,
86 	/* Direct submission to the ring buffer during init and reset. */
87 	AMDGPU_IB_POOL_DIRECT,
88 
89 	AMDGPU_IB_POOL_MAX
90 };
91 
92 struct amdgpu_ib {
93 	struct amdgpu_sa_bo		*sa_bo;
94 	uint32_t			length_dw;
95 	uint64_t			gpu_addr;
96 	uint32_t			*ptr;
97 	uint32_t			flags;
98 };
99 
100 struct amdgpu_sched {
101 	u32				num_scheds;
102 	struct drm_gpu_scheduler	*sched[AMDGPU_MAX_HWIP_RINGS];
103 };
104 
105 /*
106  * Fences.
107  */
108 struct amdgpu_fence_driver {
109 	uint64_t			gpu_addr;
110 	volatile uint32_t		*cpu_addr;
111 	/* sync_seq is protected by ring emission lock */
112 	uint32_t			sync_seq;
113 	atomic_t			last_seq;
114 	bool				initialized;
115 	struct amdgpu_irq_src		*irq_src;
116 	unsigned			irq_type;
117 	struct timer_list		fallback_timer;
118 	unsigned			num_fences_mask;
119 	spinlock_t			lock;
120 	struct dma_fence		**fences;
121 };
122 
123 extern const struct drm_sched_backend_ops amdgpu_sched_ops;
124 
125 void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring);
126 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring);
127 
128 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring);
129 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
130 				   struct amdgpu_irq_src *irq_src,
131 				   unsigned irq_type);
132 void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev);
133 void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev);
134 int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev);
135 void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev);
136 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence, struct amdgpu_job *job,
137 		      unsigned flags);
138 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
139 			      uint32_t timeout);
140 bool amdgpu_fence_process(struct amdgpu_ring *ring);
141 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);
142 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
143 				      uint32_t wait_seq,
144 				      signed long timeout);
145 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
146 
147 /*
148  * Rings.
149  */
150 
151 /* provided by hw blocks that expose a ring buffer for commands */
152 struct amdgpu_ring_funcs {
153 	enum amdgpu_ring_type	type;
154 	uint32_t		align_mask;
155 	u32			nop;
156 	bool			support_64bit_ptrs;
157 	bool			no_user_fence;
158 	bool			secure_submission_supported;
159 	unsigned		vmhub;
160 	unsigned		extra_dw;
161 
162 	/* ring read/write ptr handling */
163 	u64 (*get_rptr)(struct amdgpu_ring *ring);
164 	u64 (*get_wptr)(struct amdgpu_ring *ring);
165 	void (*set_wptr)(struct amdgpu_ring *ring);
166 	/* validating and patching of IBs */
167 	int (*parse_cs)(struct amdgpu_cs_parser *p,
168 			struct amdgpu_job *job,
169 			struct amdgpu_ib *ib);
170 	int (*patch_cs_in_place)(struct amdgpu_cs_parser *p,
171 				 struct amdgpu_job *job,
172 				 struct amdgpu_ib *ib);
173 	/* constants to calculate how many DW are needed for an emit */
174 	unsigned emit_frame_size;
175 	unsigned emit_ib_size;
176 	/* command emit functions */
177 	void (*emit_ib)(struct amdgpu_ring *ring,
178 			struct amdgpu_job *job,
179 			struct amdgpu_ib *ib,
180 			uint32_t flags);
181 	void (*emit_fence)(struct amdgpu_ring *ring, uint64_t addr,
182 			   uint64_t seq, unsigned flags);
183 	void (*emit_pipeline_sync)(struct amdgpu_ring *ring);
184 	void (*emit_vm_flush)(struct amdgpu_ring *ring, unsigned vmid,
185 			      uint64_t pd_addr);
186 	void (*emit_hdp_flush)(struct amdgpu_ring *ring);
187 	void (*emit_gds_switch)(struct amdgpu_ring *ring, uint32_t vmid,
188 				uint32_t gds_base, uint32_t gds_size,
189 				uint32_t gws_base, uint32_t gws_size,
190 				uint32_t oa_base, uint32_t oa_size);
191 	/* testing functions */
192 	int (*test_ring)(struct amdgpu_ring *ring);
193 	int (*test_ib)(struct amdgpu_ring *ring, long timeout);
194 	/* insert NOP packets */
195 	void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count);
196 	void (*insert_start)(struct amdgpu_ring *ring);
197 	void (*insert_end)(struct amdgpu_ring *ring);
198 	/* pad the indirect buffer to the necessary number of dw */
199 	void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
200 	unsigned (*init_cond_exec)(struct amdgpu_ring *ring);
201 	void (*patch_cond_exec)(struct amdgpu_ring *ring, unsigned offset);
202 	/* note usage for clock and power gating */
203 	void (*begin_use)(struct amdgpu_ring *ring);
204 	void (*end_use)(struct amdgpu_ring *ring);
205 	void (*emit_switch_buffer) (struct amdgpu_ring *ring);
206 	void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t flags);
207 	void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg,
208 			  uint32_t reg_val_offs);
209 	void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val);
210 	void (*emit_reg_wait)(struct amdgpu_ring *ring, uint32_t reg,
211 			      uint32_t val, uint32_t mask);
212 	void (*emit_reg_write_reg_wait)(struct amdgpu_ring *ring,
213 					uint32_t reg0, uint32_t reg1,
214 					uint32_t ref, uint32_t mask);
215 	void (*emit_frame_cntl)(struct amdgpu_ring *ring, bool start,
216 				bool secure);
217 	/* Try to soft recover the ring to make the fence signal */
218 	void (*soft_recovery)(struct amdgpu_ring *ring, unsigned vmid);
219 	int (*preempt_ib)(struct amdgpu_ring *ring);
220 	void (*emit_mem_sync)(struct amdgpu_ring *ring);
221 	void (*emit_wave_limit)(struct amdgpu_ring *ring, bool enable);
222 };
223 
224 struct amdgpu_ring {
225 	struct amdgpu_device		*adev;
226 	const struct amdgpu_ring_funcs	*funcs;
227 	struct amdgpu_fence_driver	fence_drv;
228 	struct drm_gpu_scheduler	sched;
229 
230 	struct amdgpu_bo	*ring_obj;
231 	volatile uint32_t	*ring;
232 	unsigned		rptr_offs;
233 	u64			rptr_gpu_addr;
234 	volatile u32		*rptr_cpu_addr;
235 	u64			wptr;
236 	u64			wptr_old;
237 	unsigned		ring_size;
238 	unsigned		max_dw;
239 	int			count_dw;
240 	uint64_t		gpu_addr;
241 	uint64_t		ptr_mask;
242 	uint32_t		buf_mask;
243 	u32			idx;
244 	u32			me;
245 	u32			pipe;
246 	u32			queue;
247 	struct amdgpu_bo	*mqd_obj;
248 	uint64_t                mqd_gpu_addr;
249 	void                    *mqd_ptr;
250 	uint64_t                eop_gpu_addr;
251 	u32			doorbell_index;
252 	bool			use_doorbell;
253 	bool			use_pollmem;
254 	unsigned		wptr_offs;
255 	u64			wptr_gpu_addr;
256 	volatile u32		*wptr_cpu_addr;
257 	unsigned		fence_offs;
258 	u64			fence_gpu_addr;
259 	volatile u32		*fence_cpu_addr;
260 	uint64_t		current_ctx;
261 	char			name[16];
262 	u32                     trail_seq;
263 	unsigned		trail_fence_offs;
264 	u64			trail_fence_gpu_addr;
265 	volatile u32		*trail_fence_cpu_addr;
266 	unsigned		cond_exe_offs;
267 	u64			cond_exe_gpu_addr;
268 	volatile u32		*cond_exe_cpu_addr;
269 	unsigned		vm_inv_eng;
270 	struct dma_fence	*vmid_wait;
271 	bool			has_compute_vm_bug;
272 	bool			no_scheduler;
273 	int			hw_prio;
274 	unsigned 		num_hw_submission;
275 	atomic_t		*sched_score;
276 
277 	/* used for mes */
278 	bool			is_mes_queue;
279 	uint32_t		hw_queue_id;
280 	struct amdgpu_mes_ctx_data *mes_ctx;
281 };
282 
283 #define amdgpu_ring_parse_cs(r, p, job, ib) ((r)->funcs->parse_cs((p), (job), (ib)))
284 #define amdgpu_ring_patch_cs_in_place(r, p, job, ib) ((r)->funcs->patch_cs_in_place((p), (job), (ib)))
285 #define amdgpu_ring_test_ring(r) (r)->funcs->test_ring((r))
286 #define amdgpu_ring_test_ib(r, t) (r)->funcs->test_ib((r), (t))
287 #define amdgpu_ring_get_rptr(r) (r)->funcs->get_rptr((r))
288 #define amdgpu_ring_get_wptr(r) (r)->funcs->get_wptr((r))
289 #define amdgpu_ring_set_wptr(r) (r)->funcs->set_wptr((r))
290 #define amdgpu_ring_emit_ib(r, job, ib, flags) ((r)->funcs->emit_ib((r), (job), (ib), (flags)))
291 #define amdgpu_ring_emit_pipeline_sync(r) (r)->funcs->emit_pipeline_sync((r))
292 #define amdgpu_ring_emit_vm_flush(r, vmid, addr) (r)->funcs->emit_vm_flush((r), (vmid), (addr))
293 #define amdgpu_ring_emit_fence(r, addr, seq, flags) (r)->funcs->emit_fence((r), (addr), (seq), (flags))
294 #define amdgpu_ring_emit_gds_switch(r, v, db, ds, wb, ws, ab, as) (r)->funcs->emit_gds_switch((r), (v), (db), (ds), (wb), (ws), (ab), (as))
295 #define amdgpu_ring_emit_hdp_flush(r) (r)->funcs->emit_hdp_flush((r))
296 #define amdgpu_ring_emit_switch_buffer(r) (r)->funcs->emit_switch_buffer((r))
297 #define amdgpu_ring_emit_cntxcntl(r, d) (r)->funcs->emit_cntxcntl((r), (d))
298 #define amdgpu_ring_emit_rreg(r, d, o) (r)->funcs->emit_rreg((r), (d), (o))
299 #define amdgpu_ring_emit_wreg(r, d, v) (r)->funcs->emit_wreg((r), (d), (v))
300 #define amdgpu_ring_emit_reg_wait(r, d, v, m) (r)->funcs->emit_reg_wait((r), (d), (v), (m))
301 #define amdgpu_ring_emit_reg_write_reg_wait(r, d0, d1, v, m) (r)->funcs->emit_reg_write_reg_wait((r), (d0), (d1), (v), (m))
302 #define amdgpu_ring_emit_frame_cntl(r, b, s) (r)->funcs->emit_frame_cntl((r), (b), (s))
303 #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
304 #define amdgpu_ring_init_cond_exec(r) (r)->funcs->init_cond_exec((r))
305 #define amdgpu_ring_patch_cond_exec(r,o) (r)->funcs->patch_cond_exec((r),(o))
306 #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r)
307 
308 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
309 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
310 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
311 void amdgpu_ring_commit(struct amdgpu_ring *ring);
312 void amdgpu_ring_undo(struct amdgpu_ring *ring);
313 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
314 		     unsigned int max_dw, struct amdgpu_irq_src *irq_src,
315 		     unsigned int irq_type, unsigned int hw_prio,
316 		     atomic_t *sched_score);
317 void amdgpu_ring_fini(struct amdgpu_ring *ring);
318 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
319 						uint32_t reg0, uint32_t val0,
320 						uint32_t reg1, uint32_t val1);
321 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
322 			       struct dma_fence *fence);
323 
324 static inline void amdgpu_ring_set_preempt_cond_exec(struct amdgpu_ring *ring,
325 							bool cond_exec)
326 {
327 	*ring->cond_exe_cpu_addr = cond_exec;
328 }
329 
330 static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
331 {
332 	int i = 0;
333 	while (i <= ring->buf_mask)
334 		ring->ring[i++] = ring->funcs->nop;
335 
336 }
337 
338 static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
339 {
340 	if (ring->count_dw <= 0)
341 		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
342 	ring->ring[ring->wptr++ & ring->buf_mask] = v;
343 	ring->wptr &= ring->ptr_mask;
344 	ring->count_dw--;
345 }
346 
347 static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
348 					      void *src, int count_dw)
349 {
350 	unsigned occupied, chunk1, chunk2;
351 	void *dst;
352 
353 	if (unlikely(ring->count_dw < count_dw))
354 		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
355 
356 	occupied = ring->wptr & ring->buf_mask;
357 	dst = (void *)&ring->ring[occupied];
358 	chunk1 = ring->buf_mask + 1 - occupied;
359 	chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
360 	chunk2 = count_dw - chunk1;
361 	chunk1 <<= 2;
362 	chunk2 <<= 2;
363 
364 	if (chunk1)
365 		memcpy(dst, src, chunk1);
366 
367 	if (chunk2) {
368 		src += chunk1;
369 		dst = (void *)ring->ring;
370 		memcpy(dst, src, chunk2);
371 	}
372 
373 	ring->wptr += count_dw;
374 	ring->wptr &= ring->ptr_mask;
375 	ring->count_dw -= count_dw;
376 }
377 
378 #define amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset)			\
379 	(ring->is_mes_queue && ring->mes_ctx ?				\
380 	 (ring->mes_ctx->meta_data_gpu_addr + offset) : 0)
381 
382 #define amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset)			\
383 	(ring->is_mes_queue && ring->mes_ctx ?				\
384 	 (void *)((uint8_t *)(ring->mes_ctx->meta_data_ptr) + offset) : \
385 	 NULL)
386 
387 int amdgpu_ring_test_helper(struct amdgpu_ring *ring);
388 
389 void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
390 			      struct amdgpu_ring *ring);
391 
392 int amdgpu_ring_init_mqd(struct amdgpu_ring *ring);
393 
394 static inline u32 amdgpu_ib_get_value(struct amdgpu_ib *ib, int idx)
395 {
396 	return ib->ptr[idx];
397 }
398 
399 static inline void amdgpu_ib_set_value(struct amdgpu_ib *ib, int idx,
400 				       uint32_t value)
401 {
402 	ib->ptr[idx] = value;
403 }
404 
405 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
406 		  unsigned size,
407 		  enum amdgpu_ib_pool_type pool,
408 		  struct amdgpu_ib *ib);
409 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
410 		    struct dma_fence *f);
411 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
412 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
413 		       struct dma_fence **f);
414 int amdgpu_ib_pool_init(struct amdgpu_device *adev);
415 void amdgpu_ib_pool_fini(struct amdgpu_device *adev);
416 int amdgpu_ib_ring_tests(struct amdgpu_device *adev);
417 
418 #endif
419