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