xref: /linux/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h (revision 07fdad3a93756b872da7b53647715c48d0f4a2d0)
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 #include <drm/drm_suballoc.h>
31 
32 struct amdgpu_device;
33 struct amdgpu_ring;
34 struct amdgpu_ib;
35 struct amdgpu_cs_parser;
36 struct amdgpu_job;
37 struct amdgpu_vm;
38 
39 /* max number of rings */
40 #define AMDGPU_MAX_RINGS		149
41 #define AMDGPU_MAX_HWIP_RINGS		64
42 #define AMDGPU_MAX_GFX_RINGS		2
43 #define AMDGPU_MAX_SW_GFX_RINGS         2
44 #define AMDGPU_MAX_COMPUTE_RINGS	8
45 #define AMDGPU_MAX_VCE_RINGS		3
46 #define AMDGPU_MAX_UVD_ENC_RINGS	2
47 #define AMDGPU_MAX_VPE_RINGS		2
48 
49 enum amdgpu_ring_priority_level {
50 	AMDGPU_RING_PRIO_0,
51 	AMDGPU_RING_PRIO_1,
52 	AMDGPU_RING_PRIO_DEFAULT = 1,
53 	AMDGPU_RING_PRIO_2,
54 	AMDGPU_RING_PRIO_MAX
55 };
56 
57 /* some special values for the owner field */
58 #define AMDGPU_FENCE_OWNER_UNDEFINED	((void *)0ul)
59 #define AMDGPU_FENCE_OWNER_VM		((void *)1ul)
60 #define AMDGPU_FENCE_OWNER_KFD		((void *)2ul)
61 
62 #define AMDGPU_FENCE_FLAG_64BIT         (1 << 0)
63 #define AMDGPU_FENCE_FLAG_INT           (1 << 1)
64 #define AMDGPU_FENCE_FLAG_TC_WB_ONLY    (1 << 2)
65 #define AMDGPU_FENCE_FLAG_EXEC          (1 << 3)
66 
67 #define to_amdgpu_ring(s) container_of((s), struct amdgpu_ring, sched)
68 
69 #define AMDGPU_IB_POOL_SIZE	(1024 * 1024)
70 
71 enum amdgpu_ring_type {
72 	AMDGPU_RING_TYPE_GFX		= AMDGPU_HW_IP_GFX,
73 	AMDGPU_RING_TYPE_COMPUTE	= AMDGPU_HW_IP_COMPUTE,
74 	AMDGPU_RING_TYPE_SDMA		= AMDGPU_HW_IP_DMA,
75 	AMDGPU_RING_TYPE_UVD		= AMDGPU_HW_IP_UVD,
76 	AMDGPU_RING_TYPE_VCE		= AMDGPU_HW_IP_VCE,
77 	AMDGPU_RING_TYPE_UVD_ENC	= AMDGPU_HW_IP_UVD_ENC,
78 	AMDGPU_RING_TYPE_VCN_DEC	= AMDGPU_HW_IP_VCN_DEC,
79 	AMDGPU_RING_TYPE_VCN_ENC	= AMDGPU_HW_IP_VCN_ENC,
80 	AMDGPU_RING_TYPE_VCN_JPEG	= AMDGPU_HW_IP_VCN_JPEG,
81 	AMDGPU_RING_TYPE_VPE		= AMDGPU_HW_IP_VPE,
82 	AMDGPU_RING_TYPE_KIQ,
83 	AMDGPU_RING_TYPE_MES,
84 	AMDGPU_RING_TYPE_UMSCH_MM,
85 	AMDGPU_RING_TYPE_CPER,
86 };
87 
88 enum amdgpu_ib_pool_type {
89 	/* Normal submissions to the top of the pipeline. */
90 	AMDGPU_IB_POOL_DELAYED,
91 	/* Immediate submissions to the bottom of the pipeline. */
92 	AMDGPU_IB_POOL_IMMEDIATE,
93 	/* Direct submission to the ring buffer during init and reset. */
94 	AMDGPU_IB_POOL_DIRECT,
95 
96 	AMDGPU_IB_POOL_MAX
97 };
98 
99 struct amdgpu_ib {
100 	struct drm_suballoc		*sa_bo;
101 	uint32_t			length_dw;
102 	uint64_t			gpu_addr;
103 	uint32_t			*ptr;
104 	uint32_t			flags;
105 };
106 
107 struct amdgpu_sched {
108 	u32				num_scheds;
109 	struct drm_gpu_scheduler	*sched[AMDGPU_MAX_HWIP_RINGS];
110 };
111 
112 /*
113  * Fences.
114  */
115 struct amdgpu_fence_driver {
116 	uint64_t			gpu_addr;
117 	uint32_t			*cpu_addr;
118 	/* sync_seq is protected by ring emission lock */
119 	uint32_t			sync_seq;
120 	atomic_t			last_seq;
121 	u64				signalled_wptr;
122 	bool				initialized;
123 	struct amdgpu_irq_src		*irq_src;
124 	unsigned			irq_type;
125 	struct timer_list		fallback_timer;
126 	unsigned			num_fences_mask;
127 	spinlock_t			lock;
128 	struct dma_fence		**fences;
129 };
130 
131 /*
132  * Fences mark an event in the GPUs pipeline and are used
133  * for GPU/CPU synchronization.  When the fence is written,
134  * it is expected that all buffers associated with that fence
135  * are no longer in use by the associated ring on the GPU and
136  * that the relevant GPU caches have been flushed.
137  */
138 
139 struct amdgpu_fence {
140 	struct dma_fence base;
141 
142 	/* RB, DMA, etc. */
143 	struct amdgpu_ring		*ring;
144 	ktime_t				start_timestamp;
145 
146 	/* wptr for the fence for resets */
147 	u64				wptr;
148 	/* fence context for resets */
149 	u64				context;
150 	uint32_t			seq;
151 };
152 
153 extern const struct drm_sched_backend_ops amdgpu_sched_ops;
154 
155 void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring);
156 void amdgpu_fence_driver_set_error(struct amdgpu_ring *ring, int error);
157 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring);
158 void amdgpu_fence_driver_guilty_force_completion(struct amdgpu_fence *fence);
159 void amdgpu_fence_save_wptr(struct dma_fence *fence);
160 
161 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring);
162 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
163 				   struct amdgpu_irq_src *irq_src,
164 				   unsigned irq_type);
165 void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev);
166 void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev);
167 int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev);
168 void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev);
169 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **f,
170 		      struct amdgpu_fence *af, unsigned int flags);
171 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
172 			      uint32_t timeout);
173 bool amdgpu_fence_process(struct amdgpu_ring *ring);
174 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);
175 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
176 				      uint32_t wait_seq,
177 				      signed long timeout);
178 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
179 
180 void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop);
181 
182 u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring);
183 void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq,
184 					 ktime_t timestamp);
185 
186 /*
187  * Rings.
188  */
189 
190 /* provided by hw blocks that expose a ring buffer for commands */
191 struct amdgpu_ring_funcs {
192 	/**
193 	 * @type:
194 	 *
195 	 * GFX, Compute, SDMA, UVD, VCE, VCN, VPE, KIQ, MES, UMSCH, and CPER
196 	 * use ring buffers. The type field just identifies which component the
197 	 * ring buffer is associated with.
198 	 */
199 	enum amdgpu_ring_type	type;
200 	uint32_t		align_mask;
201 
202 	/**
203 	 * @nop:
204 	 *
205 	 * Every block in the amdgpu has no-op instructions (e.g., GFX 10
206 	 * uses PACKET3(PACKET3_NOP, 0x3FFF), VCN 5 uses VCN_ENC_CMD_NO_OP,
207 	 * etc). This field receives the specific no-op for the component
208 	 * that initializes the ring.
209 	 */
210 	u32			nop;
211 	bool			support_64bit_ptrs;
212 	bool			no_user_fence;
213 	bool			secure_submission_supported;
214 
215 	/**
216 	 * @extra_bytes:
217 	 *
218 	 * Optional extra space in bytes that is added to the ring size
219 	 * when allocating the BO that holds the contents of the ring.
220 	 * This space isn't used for command submission to the ring,
221 	 * but is just there to satisfy some hardware requirements or
222 	 * implement workarounds. It's up to the implementation of each
223 	 * specific ring to initialize this space.
224 	 */
225 	unsigned		extra_bytes;
226 
227 	/* ring read/write ptr handling */
228 	u64 (*get_rptr)(struct amdgpu_ring *ring);
229 	u64 (*get_wptr)(struct amdgpu_ring *ring);
230 	void (*set_wptr)(struct amdgpu_ring *ring);
231 	/* validating and patching of IBs */
232 	int (*parse_cs)(struct amdgpu_cs_parser *p,
233 			struct amdgpu_job *job,
234 			struct amdgpu_ib *ib);
235 	int (*patch_cs_in_place)(struct amdgpu_cs_parser *p,
236 				 struct amdgpu_job *job,
237 				 struct amdgpu_ib *ib);
238 	/* constants to calculate how many DW are needed for an emit */
239 	unsigned emit_frame_size;
240 	unsigned emit_ib_size;
241 	/* command emit functions */
242 	void (*emit_ib)(struct amdgpu_ring *ring,
243 			struct amdgpu_job *job,
244 			struct amdgpu_ib *ib,
245 			uint32_t flags);
246 	void (*emit_fence)(struct amdgpu_ring *ring, uint64_t addr,
247 			   uint64_t seq, unsigned flags);
248 	void (*emit_pipeline_sync)(struct amdgpu_ring *ring);
249 	void (*emit_vm_flush)(struct amdgpu_ring *ring, unsigned vmid,
250 			      uint64_t pd_addr);
251 	void (*emit_hdp_flush)(struct amdgpu_ring *ring);
252 	void (*emit_gds_switch)(struct amdgpu_ring *ring, uint32_t vmid,
253 				uint32_t gds_base, uint32_t gds_size,
254 				uint32_t gws_base, uint32_t gws_size,
255 				uint32_t oa_base, uint32_t oa_size);
256 	/* testing functions */
257 	int (*test_ring)(struct amdgpu_ring *ring);
258 	int (*test_ib)(struct amdgpu_ring *ring, long timeout);
259 	/* insert NOP packets */
260 	void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count);
261 	void (*insert_start)(struct amdgpu_ring *ring);
262 	void (*insert_end)(struct amdgpu_ring *ring);
263 	/* pad the indirect buffer to the necessary number of dw */
264 	void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
265 	unsigned (*init_cond_exec)(struct amdgpu_ring *ring, uint64_t addr);
266 	/* note usage for clock and power gating */
267 	void (*begin_use)(struct amdgpu_ring *ring);
268 	void (*end_use)(struct amdgpu_ring *ring);
269 	void (*emit_switch_buffer) (struct amdgpu_ring *ring);
270 	void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t flags);
271 	void (*emit_gfx_shadow)(struct amdgpu_ring *ring, u64 shadow_va, u64 csa_va,
272 				u64 gds_va, bool init_shadow, int vmid);
273 	void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg,
274 			  uint32_t reg_val_offs);
275 	void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val);
276 	void (*emit_reg_wait)(struct amdgpu_ring *ring, uint32_t reg,
277 			      uint32_t val, uint32_t mask);
278 	void (*emit_reg_write_reg_wait)(struct amdgpu_ring *ring,
279 					uint32_t reg0, uint32_t reg1,
280 					uint32_t ref, uint32_t mask);
281 	void (*emit_frame_cntl)(struct amdgpu_ring *ring, bool start,
282 				bool secure);
283 	/* Try to soft recover the ring to make the fence signal */
284 	void (*soft_recovery)(struct amdgpu_ring *ring, unsigned vmid);
285 	int (*preempt_ib)(struct amdgpu_ring *ring);
286 	void (*emit_mem_sync)(struct amdgpu_ring *ring);
287 	void (*emit_wave_limit)(struct amdgpu_ring *ring, bool enable);
288 	void (*patch_cntl)(struct amdgpu_ring *ring, unsigned offset);
289 	void (*patch_ce)(struct amdgpu_ring *ring, unsigned offset);
290 	void (*patch_de)(struct amdgpu_ring *ring, unsigned offset);
291 	int (*reset)(struct amdgpu_ring *ring, unsigned int vmid,
292 		     struct amdgpu_fence *timedout_fence);
293 	void (*emit_cleaner_shader)(struct amdgpu_ring *ring);
294 };
295 
296 /**
297  * amdgpu_ring - Holds ring information
298  */
299 struct amdgpu_ring {
300 	struct amdgpu_device		*adev;
301 	const struct amdgpu_ring_funcs	*funcs;
302 	struct amdgpu_fence_driver	fence_drv;
303 	struct drm_gpu_scheduler	sched;
304 
305 	struct amdgpu_bo	*ring_obj;
306 	uint32_t		*ring;
307 	/* backups for resets */
308 	uint32_t		*ring_backup;
309 	unsigned int		ring_backup_entries_to_copy;
310 	unsigned		rptr_offs;
311 	u64			rptr_gpu_addr;
312 	u32			*rptr_cpu_addr;
313 
314 	/**
315 	 * @wptr:
316 	 *
317 	 * This is part of the Ring buffer implementation and represents the
318 	 * write pointer. The wptr determines where the host has written.
319 	 */
320 	u64			wptr;
321 
322 	/**
323 	 * @wptr_old:
324 	 *
325 	 * Before update wptr with the new value, usually the old value is
326 	 * stored in the wptr_old.
327 	 */
328 	u64			wptr_old;
329 	unsigned		ring_size;
330 
331 	/**
332 	 * @max_dw:
333 	 *
334 	 * Maximum number of DWords for ring allocation. This information is
335 	 * provided at the ring initialization time, and each IP block can
336 	 * specify a specific value. Check places that invoke
337 	 * amdgpu_ring_init() to see the maximum size per block.
338 	 */
339 	unsigned		max_dw;
340 
341 	/**
342 	 * @count_dw:
343 	 *
344 	 * This value starts with the maximum amount of DWords supported by the
345 	 * ring. This value is updated based on the ring manipulation.
346 	 */
347 	int			count_dw;
348 	uint64_t		gpu_addr;
349 
350 	/**
351 	 * @ptr_mask:
352 	 *
353 	 * Some IPs provide support for 64-bit pointers and others for 32-bit
354 	 * only; this behavior is component-specific and defined by the field
355 	 * support_64bit_ptr. If the IP block supports 64-bits, the mask
356 	 * 0xffffffffffffffff is set; otherwise, this value assumes buf_mask.
357 	 * Notice that this field is used to keep wptr under a valid range.
358 	 */
359 	uint64_t		ptr_mask;
360 
361 	/**
362 	 * @buf_mask:
363 	 *
364 	 * Buffer mask is a value used to keep wptr count under its
365 	 * thresholding. Buffer mask initialized during the ring buffer
366 	 * initialization time, and it is defined as (ring_size / 4) -1.
367 	 */
368 	uint32_t		buf_mask;
369 	u32			idx;
370 	u32			xcc_id;
371 	u32			xcp_id;
372 	u32			me;
373 	u32			pipe;
374 	u32			queue;
375 	struct amdgpu_bo	*mqd_obj;
376 	uint64_t                mqd_gpu_addr;
377 	void                    *mqd_ptr;
378 	unsigned                mqd_size;
379 	uint64_t                eop_gpu_addr;
380 	u32			doorbell_index;
381 	bool			use_doorbell;
382 	bool			use_pollmem;
383 	unsigned		wptr_offs;
384 	u64			wptr_gpu_addr;
385 
386 	/**
387 	 * @wptr_cpu_addr:
388 	 *
389 	 * This is the CPU address pointer in the writeback slot. This is used
390 	 * to commit changes to the GPU.
391 	 */
392 	u32			*wptr_cpu_addr;
393 	unsigned		fence_offs;
394 	u64			fence_gpu_addr;
395 	u32			*fence_cpu_addr;
396 	uint64_t		current_ctx;
397 	char			name[16];
398 	u32                     trail_seq;
399 	unsigned		trail_fence_offs;
400 	u64			trail_fence_gpu_addr;
401 	u32			*trail_fence_cpu_addr;
402 	unsigned		cond_exe_offs;
403 	u64			cond_exe_gpu_addr;
404 	u32			*cond_exe_cpu_addr;
405 	unsigned int		set_q_mode_offs;
406 	u32			*set_q_mode_ptr;
407 	u64			set_q_mode_token;
408 	unsigned		vm_hub;
409 	unsigned		vm_inv_eng;
410 	struct dma_fence	*vmid_wait;
411 	bool			has_compute_vm_bug;
412 	bool			no_scheduler;
413 	bool			no_user_submission;
414 	int			hw_prio;
415 	unsigned 		num_hw_submission;
416 	atomic_t		*sched_score;
417 
418 	bool            is_sw_ring;
419 	unsigned int    entry_index;
420 	/* store the cached rptr to restore after reset */
421 	uint64_t cached_rptr;
422 };
423 
424 #define amdgpu_ring_parse_cs(r, p, job, ib) ((r)->funcs->parse_cs((p), (job), (ib)))
425 #define amdgpu_ring_patch_cs_in_place(r, p, job, ib) ((r)->funcs->patch_cs_in_place((p), (job), (ib)))
426 #define amdgpu_ring_test_ring(r) (r)->funcs->test_ring((r))
427 #define amdgpu_ring_test_ib(r, t) ((r)->funcs->test_ib ? (r)->funcs->test_ib((r), (t)) : 0)
428 #define amdgpu_ring_get_rptr(r) (r)->funcs->get_rptr((r))
429 #define amdgpu_ring_get_wptr(r) (r)->funcs->get_wptr((r))
430 #define amdgpu_ring_set_wptr(r) (r)->funcs->set_wptr((r))
431 #define amdgpu_ring_emit_ib(r, job, ib, flags) ((r)->funcs->emit_ib((r), (job), (ib), (flags)))
432 #define amdgpu_ring_emit_pipeline_sync(r) (r)->funcs->emit_pipeline_sync((r))
433 #define amdgpu_ring_emit_vm_flush(r, vmid, addr) (r)->funcs->emit_vm_flush((r), (vmid), (addr))
434 #define amdgpu_ring_emit_fence(r, addr, seq, flags) (r)->funcs->emit_fence((r), (addr), (seq), (flags))
435 #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))
436 #define amdgpu_ring_emit_hdp_flush(r) (r)->funcs->emit_hdp_flush((r))
437 #define amdgpu_ring_emit_switch_buffer(r) (r)->funcs->emit_switch_buffer((r))
438 #define amdgpu_ring_emit_cntxcntl(r, d) (r)->funcs->emit_cntxcntl((r), (d))
439 #define amdgpu_ring_emit_gfx_shadow(r, s, c, g, i, v) ((r)->funcs->emit_gfx_shadow((r), (s), (c), (g), (i), (v)))
440 #define amdgpu_ring_emit_rreg(r, d, o) (r)->funcs->emit_rreg((r), (d), (o))
441 #define amdgpu_ring_emit_wreg(r, d, v) (r)->funcs->emit_wreg((r), (d), (v))
442 #define amdgpu_ring_emit_reg_wait(r, d, v, m) (r)->funcs->emit_reg_wait((r), (d), (v), (m))
443 #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))
444 #define amdgpu_ring_emit_frame_cntl(r, b, s) (r)->funcs->emit_frame_cntl((r), (b), (s))
445 #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
446 #define amdgpu_ring_init_cond_exec(r, a) (r)->funcs->init_cond_exec((r), (a))
447 #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r)
448 #define amdgpu_ring_patch_cntl(r, o) ((r)->funcs->patch_cntl((r), (o)))
449 #define amdgpu_ring_patch_ce(r, o) ((r)->funcs->patch_ce((r), (o)))
450 #define amdgpu_ring_patch_de(r, o) ((r)->funcs->patch_de((r), (o)))
451 #define amdgpu_ring_reset(r, v, f) (r)->funcs->reset((r), (v), (f))
452 
453 unsigned int amdgpu_ring_max_ibs(enum amdgpu_ring_type type);
454 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
455 void amdgpu_ring_ib_begin(struct amdgpu_ring *ring);
456 void amdgpu_ring_ib_end(struct amdgpu_ring *ring);
457 void amdgpu_ring_ib_on_emit_cntl(struct amdgpu_ring *ring);
458 void amdgpu_ring_ib_on_emit_ce(struct amdgpu_ring *ring);
459 void amdgpu_ring_ib_on_emit_de(struct amdgpu_ring *ring);
460 
461 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
462 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
463 void amdgpu_ring_commit(struct amdgpu_ring *ring);
464 void amdgpu_ring_undo(struct amdgpu_ring *ring);
465 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
466 		     unsigned int max_dw, struct amdgpu_irq_src *irq_src,
467 		     unsigned int irq_type, unsigned int hw_prio,
468 		     atomic_t *sched_score);
469 void amdgpu_ring_fini(struct amdgpu_ring *ring);
470 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
471 						uint32_t reg0, uint32_t val0,
472 						uint32_t reg1, uint32_t val1);
473 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
474 			       struct dma_fence *fence);
475 
476 static inline void amdgpu_ring_set_preempt_cond_exec(struct amdgpu_ring *ring,
477 							bool cond_exec)
478 {
479 	*ring->cond_exe_cpu_addr = cond_exec;
480 }
481 
482 static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
483 {
484 	memset32(ring->ring, ring->funcs->nop, ring->buf_mask + 1);
485 }
486 
487 static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
488 {
489 	ring->ring[ring->wptr++ & ring->buf_mask] = v;
490 	ring->wptr &= ring->ptr_mask;
491 	ring->count_dw--;
492 }
493 
494 static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
495 					      void *src, int count_dw)
496 {
497 	unsigned occupied, chunk1, chunk2;
498 
499 	occupied = ring->wptr & ring->buf_mask;
500 	chunk1 = ring->buf_mask + 1 - occupied;
501 	chunk1 = (chunk1 >= count_dw) ? count_dw : chunk1;
502 	chunk2 = count_dw - chunk1;
503 	chunk1 <<= 2;
504 	chunk2 <<= 2;
505 
506 	if (chunk1)
507 		memcpy(&ring->ring[occupied], src, chunk1);
508 
509 	if (chunk2) {
510 		src += chunk1;
511 		memcpy(ring->ring, src, chunk2);
512 	}
513 
514 	ring->wptr += count_dw;
515 	ring->wptr &= ring->ptr_mask;
516 	ring->count_dw -= count_dw;
517 }
518 
519 /**
520  * amdgpu_ring_patch_cond_exec - patch dw count of conditional execute
521  * @ring: amdgpu_ring structure
522  * @offset: offset returned by amdgpu_ring_init_cond_exec
523  *
524  * Calculate the dw count and patch it into a cond_exec command.
525  */
526 static inline void amdgpu_ring_patch_cond_exec(struct amdgpu_ring *ring,
527 					       unsigned int offset)
528 {
529 	unsigned cur;
530 
531 	if (!ring->funcs->init_cond_exec)
532 		return;
533 
534 	WARN_ON(offset > ring->buf_mask);
535 	WARN_ON(ring->ring[offset] != 0);
536 
537 	cur = (ring->wptr - 1) & ring->buf_mask;
538 	if (cur < offset)
539 		cur += ring->ring_size >> 2;
540 	ring->ring[offset] = cur - offset;
541 }
542 
543 int amdgpu_ring_test_helper(struct amdgpu_ring *ring);
544 
545 void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
546 			      struct amdgpu_ring *ring);
547 
548 int amdgpu_ring_init_mqd(struct amdgpu_ring *ring);
549 
550 static inline u32 amdgpu_ib_get_value(struct amdgpu_ib *ib, int idx)
551 {
552 	return ib->ptr[idx];
553 }
554 
555 static inline void amdgpu_ib_set_value(struct amdgpu_ib *ib, int idx,
556 				       uint32_t value)
557 {
558 	ib->ptr[idx] = value;
559 }
560 
561 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
562 		  unsigned size,
563 		  enum amdgpu_ib_pool_type pool,
564 		  struct amdgpu_ib *ib);
565 void amdgpu_ib_free(struct amdgpu_ib *ib, struct dma_fence *f);
566 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
567 		       struct amdgpu_ib *ibs, struct amdgpu_job *job,
568 		       struct dma_fence **f);
569 int amdgpu_ib_pool_init(struct amdgpu_device *adev);
570 void amdgpu_ib_pool_fini(struct amdgpu_device *adev);
571 int amdgpu_ib_ring_tests(struct amdgpu_device *adev);
572 bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring);
573 void amdgpu_ring_backup_unprocessed_commands(struct amdgpu_ring *ring,
574 					     struct amdgpu_fence *guilty_fence);
575 void amdgpu_ring_reset_helper_begin(struct amdgpu_ring *ring,
576 				    struct amdgpu_fence *guilty_fence);
577 int amdgpu_ring_reset_helper_end(struct amdgpu_ring *ring,
578 				 struct amdgpu_fence *guilty_fence);
579 bool amdgpu_ring_is_reset_type_supported(struct amdgpu_ring *ring,
580 					 u32 reset_type);
581 #endif
582