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		24
33 #define AMDGPU_MAX_GFX_RINGS		2
34 #define AMDGPU_MAX_COMPUTE_RINGS	8
35 #define AMDGPU_MAX_VCE_RINGS		3
36 #define AMDGPU_MAX_UVD_ENC_RINGS	2
37 
38 /* some special values for the owner field */
39 #define AMDGPU_FENCE_OWNER_UNDEFINED	((void *)0ul)
40 #define AMDGPU_FENCE_OWNER_VM		((void *)1ul)
41 #define AMDGPU_FENCE_OWNER_KFD		((void *)2ul)
42 
43 #define AMDGPU_FENCE_FLAG_64BIT         (1 << 0)
44 #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
45 #define AMDGPU_FENCE_FLAG_TC_WB_ONLY    (1 << 2)
46 
47 #define to_amdgpu_ring(s) container_of((s), struct amdgpu_ring, sched)
48 
49 enum amdgpu_ring_type {
50 	AMDGPU_RING_TYPE_GFX,
51 	AMDGPU_RING_TYPE_COMPUTE,
52 	AMDGPU_RING_TYPE_SDMA,
53 	AMDGPU_RING_TYPE_UVD,
54 	AMDGPU_RING_TYPE_VCE,
55 	AMDGPU_RING_TYPE_KIQ,
56 	AMDGPU_RING_TYPE_UVD_ENC,
57 	AMDGPU_RING_TYPE_VCN_DEC,
58 	AMDGPU_RING_TYPE_VCN_ENC,
59 	AMDGPU_RING_TYPE_VCN_JPEG
60 };
61 
62 struct amdgpu_device;
63 struct amdgpu_ring;
64 struct amdgpu_ib;
65 struct amdgpu_cs_parser;
66 struct amdgpu_job;
67 
68 /*
69  * Fences.
70  */
71 struct amdgpu_fence_driver {
72 	uint64_t			gpu_addr;
73 	volatile uint32_t		*cpu_addr;
74 	/* sync_seq is protected by ring emission lock */
75 	uint32_t			sync_seq;
76 	atomic_t			last_seq;
77 	bool				initialized;
78 	struct amdgpu_irq_src		*irq_src;
79 	unsigned			irq_type;
80 	struct timer_list		fallback_timer;
81 	unsigned			num_fences_mask;
82 	spinlock_t			lock;
83 	struct dma_fence		**fences;
84 };
85 
86 int amdgpu_fence_driver_init(struct amdgpu_device *adev);
87 void amdgpu_fence_driver_fini(struct amdgpu_device *adev);
88 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring);
89 
90 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring,
91 				  unsigned num_hw_submission);
92 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
93 				   struct amdgpu_irq_src *irq_src,
94 				   unsigned irq_type);
95 void amdgpu_fence_driver_suspend(struct amdgpu_device *adev);
96 void amdgpu_fence_driver_resume(struct amdgpu_device *adev);
97 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence,
98 		      unsigned flags);
99 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s);
100 bool amdgpu_fence_process(struct amdgpu_ring *ring);
101 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);
102 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
103 				      uint32_t wait_seq,
104 				      signed long timeout);
105 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
106 
107 /*
108  * Rings.
109  */
110 
111 /* provided by hw blocks that expose a ring buffer for commands */
112 struct amdgpu_ring_funcs {
113 	enum amdgpu_ring_type	type;
114 	uint32_t		align_mask;
115 	u32			nop;
116 	bool			support_64bit_ptrs;
117 	bool			no_user_fence;
118 	unsigned		vmhub;
119 	unsigned		extra_dw;
120 
121 	/* ring read/write ptr handling */
122 	u64 (*get_rptr)(struct amdgpu_ring *ring);
123 	u64 (*get_wptr)(struct amdgpu_ring *ring);
124 	void (*set_wptr)(struct amdgpu_ring *ring);
125 	/* validating and patching of IBs */
126 	int (*parse_cs)(struct amdgpu_cs_parser *p, uint32_t ib_idx);
127 	int (*patch_cs_in_place)(struct amdgpu_cs_parser *p, uint32_t ib_idx);
128 	/* constants to calculate how many DW are needed for an emit */
129 	unsigned emit_frame_size;
130 	unsigned emit_ib_size;
131 	/* command emit functions */
132 	void (*emit_ib)(struct amdgpu_ring *ring,
133 			struct amdgpu_job *job,
134 			struct amdgpu_ib *ib,
135 			uint32_t flags);
136 	void (*emit_fence)(struct amdgpu_ring *ring, uint64_t addr,
137 			   uint64_t seq, unsigned flags);
138 	void (*emit_pipeline_sync)(struct amdgpu_ring *ring);
139 	void (*emit_vm_flush)(struct amdgpu_ring *ring, unsigned vmid,
140 			      uint64_t pd_addr);
141 	void (*emit_hdp_flush)(struct amdgpu_ring *ring);
142 	void (*emit_gds_switch)(struct amdgpu_ring *ring, uint32_t vmid,
143 				uint32_t gds_base, uint32_t gds_size,
144 				uint32_t gws_base, uint32_t gws_size,
145 				uint32_t oa_base, uint32_t oa_size);
146 	/* testing functions */
147 	int (*test_ring)(struct amdgpu_ring *ring);
148 	int (*test_ib)(struct amdgpu_ring *ring, long timeout);
149 	/* insert NOP packets */
150 	void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count);
151 	void (*insert_start)(struct amdgpu_ring *ring);
152 	void (*insert_end)(struct amdgpu_ring *ring);
153 	/* pad the indirect buffer to the necessary number of dw */
154 	void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
155 	unsigned (*init_cond_exec)(struct amdgpu_ring *ring);
156 	void (*patch_cond_exec)(struct amdgpu_ring *ring, unsigned offset);
157 	/* note usage for clock and power gating */
158 	void (*begin_use)(struct amdgpu_ring *ring);
159 	void (*end_use)(struct amdgpu_ring *ring);
160 	void (*emit_switch_buffer) (struct amdgpu_ring *ring);
161 	void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t flags);
162 	void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg);
163 	void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val);
164 	void (*emit_reg_wait)(struct amdgpu_ring *ring, uint32_t reg,
165 			      uint32_t val, uint32_t mask);
166 	void (*emit_reg_write_reg_wait)(struct amdgpu_ring *ring,
167 					uint32_t reg0, uint32_t reg1,
168 					uint32_t ref, uint32_t mask);
169 	void (*emit_tmz)(struct amdgpu_ring *ring, bool start);
170 	/* priority functions */
171 	void (*set_priority) (struct amdgpu_ring *ring,
172 			      enum drm_sched_priority priority);
173 	/* Try to soft recover the ring to make the fence signal */
174 	void (*soft_recovery)(struct amdgpu_ring *ring, unsigned vmid);
175 	int (*preempt_ib)(struct amdgpu_ring *ring);
176 };
177 
178 struct amdgpu_ring {
179 	struct amdgpu_device		*adev;
180 	const struct amdgpu_ring_funcs	*funcs;
181 	struct amdgpu_fence_driver	fence_drv;
182 	struct drm_gpu_scheduler	sched;
183 
184 	struct amdgpu_bo	*ring_obj;
185 	volatile uint32_t	*ring;
186 	unsigned		rptr_offs;
187 	u64			wptr;
188 	u64			wptr_old;
189 	unsigned		ring_size;
190 	unsigned		max_dw;
191 	int			count_dw;
192 	uint64_t		gpu_addr;
193 	uint64_t		ptr_mask;
194 	uint32_t		buf_mask;
195 	u32			idx;
196 	u32			me;
197 	u32			pipe;
198 	u32			queue;
199 	struct amdgpu_bo	*mqd_obj;
200 	uint64_t                mqd_gpu_addr;
201 	void                    *mqd_ptr;
202 	uint64_t                eop_gpu_addr;
203 	u32			doorbell_index;
204 	bool			use_doorbell;
205 	bool			use_pollmem;
206 	unsigned		wptr_offs;
207 	unsigned		fence_offs;
208 	uint64_t		current_ctx;
209 	char			name[16];
210 	u32                     trail_seq;
211 	unsigned		trail_fence_offs;
212 	u64			trail_fence_gpu_addr;
213 	volatile u32		*trail_fence_cpu_addr;
214 	unsigned		cond_exe_offs;
215 	u64			cond_exe_gpu_addr;
216 	volatile u32		*cond_exe_cpu_addr;
217 	unsigned		vm_inv_eng;
218 	struct dma_fence	*vmid_wait;
219 	bool			has_compute_vm_bug;
220 
221 	atomic_t		num_jobs[DRM_SCHED_PRIORITY_MAX];
222 	struct mutex		priority_mutex;
223 	/* protected by priority_mutex */
224 	int			priority;
225 
226 #if defined(CONFIG_DEBUG_FS)
227 	struct dentry *ent;
228 #endif
229 };
230 
231 #define amdgpu_ring_parse_cs(r, p, ib) ((r)->funcs->parse_cs((p), (ib)))
232 #define amdgpu_ring_patch_cs_in_place(r, p, ib) ((r)->funcs->patch_cs_in_place((p), (ib)))
233 #define amdgpu_ring_test_ring(r) (r)->funcs->test_ring((r))
234 #define amdgpu_ring_test_ib(r, t) (r)->funcs->test_ib((r), (t))
235 #define amdgpu_ring_get_rptr(r) (r)->funcs->get_rptr((r))
236 #define amdgpu_ring_get_wptr(r) (r)->funcs->get_wptr((r))
237 #define amdgpu_ring_set_wptr(r) (r)->funcs->set_wptr((r))
238 #define amdgpu_ring_emit_ib(r, job, ib, flags) ((r)->funcs->emit_ib((r), (job), (ib), (flags)))
239 #define amdgpu_ring_emit_pipeline_sync(r) (r)->funcs->emit_pipeline_sync((r))
240 #define amdgpu_ring_emit_vm_flush(r, vmid, addr) (r)->funcs->emit_vm_flush((r), (vmid), (addr))
241 #define amdgpu_ring_emit_fence(r, addr, seq, flags) (r)->funcs->emit_fence((r), (addr), (seq), (flags))
242 #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))
243 #define amdgpu_ring_emit_hdp_flush(r) (r)->funcs->emit_hdp_flush((r))
244 #define amdgpu_ring_emit_switch_buffer(r) (r)->funcs->emit_switch_buffer((r))
245 #define amdgpu_ring_emit_cntxcntl(r, d) (r)->funcs->emit_cntxcntl((r), (d))
246 #define amdgpu_ring_emit_rreg(r, d) (r)->funcs->emit_rreg((r), (d))
247 #define amdgpu_ring_emit_wreg(r, d, v) (r)->funcs->emit_wreg((r), (d), (v))
248 #define amdgpu_ring_emit_reg_wait(r, d, v, m) (r)->funcs->emit_reg_wait((r), (d), (v), (m))
249 #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))
250 #define amdgpu_ring_emit_tmz(r, b) (r)->funcs->emit_tmz((r), (b))
251 #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
252 #define amdgpu_ring_init_cond_exec(r) (r)->funcs->init_cond_exec((r))
253 #define amdgpu_ring_patch_cond_exec(r,o) (r)->funcs->patch_cond_exec((r),(o))
254 #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r)
255 
256 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
257 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
258 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
259 void amdgpu_ring_commit(struct amdgpu_ring *ring);
260 void amdgpu_ring_undo(struct amdgpu_ring *ring);
261 void amdgpu_ring_priority_get(struct amdgpu_ring *ring,
262 			      enum drm_sched_priority priority);
263 void amdgpu_ring_priority_put(struct amdgpu_ring *ring,
264 			      enum drm_sched_priority priority);
265 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
266 		     unsigned ring_size, struct amdgpu_irq_src *irq_src,
267 		     unsigned irq_type);
268 void amdgpu_ring_fini(struct amdgpu_ring *ring);
269 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
270 						uint32_t reg0, uint32_t val0,
271 						uint32_t reg1, uint32_t val1);
272 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
273 			       struct dma_fence *fence);
274 
275 static inline void amdgpu_ring_set_preempt_cond_exec(struct amdgpu_ring *ring,
276 							bool cond_exec)
277 {
278 	*ring->cond_exe_cpu_addr = cond_exec;
279 }
280 
281 static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
282 {
283 	int i = 0;
284 	while (i <= ring->buf_mask)
285 		ring->ring[i++] = ring->funcs->nop;
286 
287 }
288 
289 static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
290 {
291 	if (ring->count_dw <= 0)
292 		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
293 	ring->ring[ring->wptr++ & ring->buf_mask] = v;
294 	ring->wptr &= ring->ptr_mask;
295 	ring->count_dw--;
296 }
297 
298 static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
299 					      void *src, int count_dw)
300 {
301 	unsigned occupied, chunk1, chunk2;
302 	void *dst;
303 
304 	if (unlikely(ring->count_dw < count_dw))
305 		DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
306 
307 	occupied = ring->wptr & ring->buf_mask;
308 	dst = (void *)&ring->ring[occupied];
309 	chunk1 = ring->buf_mask + 1 - occupied;
310 	chunk1 = (chunk1 >= count_dw) ? count_dw: chunk1;
311 	chunk2 = count_dw - chunk1;
312 	chunk1 <<= 2;
313 	chunk2 <<= 2;
314 
315 	if (chunk1)
316 		memcpy(dst, src, chunk1);
317 
318 	if (chunk2) {
319 		src += chunk1;
320 		dst = (void *)ring->ring;
321 		memcpy(dst, src, chunk2);
322 	}
323 
324 	ring->wptr += count_dw;
325 	ring->wptr &= ring->ptr_mask;
326 	ring->count_dw -= count_dw;
327 }
328 
329 int amdgpu_ring_test_helper(struct amdgpu_ring *ring);
330 
331 #endif
332