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 124
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 };
86
87 enum amdgpu_ib_pool_type {
88 /* Normal submissions to the top of the pipeline. */
89 AMDGPU_IB_POOL_DELAYED,
90 /* Immediate submissions to the bottom of the pipeline. */
91 AMDGPU_IB_POOL_IMMEDIATE,
92 /* Direct submission to the ring buffer during init and reset. */
93 AMDGPU_IB_POOL_DIRECT,
94
95 AMDGPU_IB_POOL_MAX
96 };
97
98 struct amdgpu_ib {
99 struct drm_suballoc *sa_bo;
100 uint32_t length_dw;
101 uint64_t gpu_addr;
102 uint32_t *ptr;
103 uint32_t flags;
104 };
105
106 struct amdgpu_sched {
107 u32 num_scheds;
108 struct drm_gpu_scheduler *sched[AMDGPU_MAX_HWIP_RINGS];
109 };
110
111 /*
112 * Fences.
113 */
114 struct amdgpu_fence_driver {
115 uint64_t gpu_addr;
116 volatile uint32_t *cpu_addr;
117 /* sync_seq is protected by ring emission lock */
118 uint32_t sync_seq;
119 atomic_t last_seq;
120 bool initialized;
121 struct amdgpu_irq_src *irq_src;
122 unsigned irq_type;
123 struct timer_list fallback_timer;
124 unsigned num_fences_mask;
125 spinlock_t lock;
126 struct dma_fence **fences;
127 };
128
129 extern const struct drm_sched_backend_ops amdgpu_sched_ops;
130
131 void amdgpu_fence_driver_clear_job_fences(struct amdgpu_ring *ring);
132 void amdgpu_fence_driver_set_error(struct amdgpu_ring *ring, int error);
133 void amdgpu_fence_driver_force_completion(struct amdgpu_ring *ring);
134
135 int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring);
136 int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
137 struct amdgpu_irq_src *irq_src,
138 unsigned irq_type);
139 void amdgpu_fence_driver_hw_init(struct amdgpu_device *adev);
140 void amdgpu_fence_driver_hw_fini(struct amdgpu_device *adev);
141 int amdgpu_fence_driver_sw_init(struct amdgpu_device *adev);
142 void amdgpu_fence_driver_sw_fini(struct amdgpu_device *adev);
143 int amdgpu_fence_emit(struct amdgpu_ring *ring, struct dma_fence **fence, struct amdgpu_job *job,
144 unsigned flags);
145 int amdgpu_fence_emit_polling(struct amdgpu_ring *ring, uint32_t *s,
146 uint32_t timeout);
147 bool amdgpu_fence_process(struct amdgpu_ring *ring);
148 int amdgpu_fence_wait_empty(struct amdgpu_ring *ring);
149 signed long amdgpu_fence_wait_polling(struct amdgpu_ring *ring,
150 uint32_t wait_seq,
151 signed long timeout);
152 unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
153
154 void amdgpu_fence_driver_isr_toggle(struct amdgpu_device *adev, bool stop);
155
156 u64 amdgpu_fence_last_unsignaled_time_us(struct amdgpu_ring *ring);
157 void amdgpu_fence_update_start_timestamp(struct amdgpu_ring *ring, uint32_t seq,
158 ktime_t timestamp);
159
160 /*
161 * Rings.
162 */
163
164 /* provided by hw blocks that expose a ring buffer for commands */
165 struct amdgpu_ring_funcs {
166 enum amdgpu_ring_type type;
167 uint32_t align_mask;
168 u32 nop;
169 bool support_64bit_ptrs;
170 bool no_user_fence;
171 bool secure_submission_supported;
172 unsigned extra_dw;
173
174 /* ring read/write ptr handling */
175 u64 (*get_rptr)(struct amdgpu_ring *ring);
176 u64 (*get_wptr)(struct amdgpu_ring *ring);
177 void (*set_wptr)(struct amdgpu_ring *ring);
178 /* validating and patching of IBs */
179 int (*parse_cs)(struct amdgpu_cs_parser *p,
180 struct amdgpu_job *job,
181 struct amdgpu_ib *ib);
182 int (*patch_cs_in_place)(struct amdgpu_cs_parser *p,
183 struct amdgpu_job *job,
184 struct amdgpu_ib *ib);
185 /* constants to calculate how many DW are needed for an emit */
186 unsigned emit_frame_size;
187 unsigned emit_ib_size;
188 /* command emit functions */
189 void (*emit_ib)(struct amdgpu_ring *ring,
190 struct amdgpu_job *job,
191 struct amdgpu_ib *ib,
192 uint32_t flags);
193 void (*emit_fence)(struct amdgpu_ring *ring, uint64_t addr,
194 uint64_t seq, unsigned flags);
195 void (*emit_pipeline_sync)(struct amdgpu_ring *ring);
196 void (*emit_vm_flush)(struct amdgpu_ring *ring, unsigned vmid,
197 uint64_t pd_addr);
198 void (*emit_hdp_flush)(struct amdgpu_ring *ring);
199 void (*emit_gds_switch)(struct amdgpu_ring *ring, uint32_t vmid,
200 uint32_t gds_base, uint32_t gds_size,
201 uint32_t gws_base, uint32_t gws_size,
202 uint32_t oa_base, uint32_t oa_size);
203 /* testing functions */
204 int (*test_ring)(struct amdgpu_ring *ring);
205 int (*test_ib)(struct amdgpu_ring *ring, long timeout);
206 /* insert NOP packets */
207 void (*insert_nop)(struct amdgpu_ring *ring, uint32_t count);
208 void (*insert_start)(struct amdgpu_ring *ring);
209 void (*insert_end)(struct amdgpu_ring *ring);
210 /* pad the indirect buffer to the necessary number of dw */
211 void (*pad_ib)(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
212 unsigned (*init_cond_exec)(struct amdgpu_ring *ring);
213 void (*patch_cond_exec)(struct amdgpu_ring *ring, unsigned offset);
214 /* note usage for clock and power gating */
215 void (*begin_use)(struct amdgpu_ring *ring);
216 void (*end_use)(struct amdgpu_ring *ring);
217 void (*emit_switch_buffer) (struct amdgpu_ring *ring);
218 void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t flags);
219 void (*emit_gfx_shadow)(struct amdgpu_ring *ring, u64 shadow_va, u64 csa_va,
220 u64 gds_va, bool init_shadow, int vmid);
221 void (*emit_rreg)(struct amdgpu_ring *ring, uint32_t reg,
222 uint32_t reg_val_offs);
223 void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg, uint32_t val);
224 void (*emit_reg_wait)(struct amdgpu_ring *ring, uint32_t reg,
225 uint32_t val, uint32_t mask);
226 void (*emit_reg_write_reg_wait)(struct amdgpu_ring *ring,
227 uint32_t reg0, uint32_t reg1,
228 uint32_t ref, uint32_t mask);
229 void (*emit_frame_cntl)(struct amdgpu_ring *ring, bool start,
230 bool secure);
231 /* Try to soft recover the ring to make the fence signal */
232 void (*soft_recovery)(struct amdgpu_ring *ring, unsigned vmid);
233 int (*preempt_ib)(struct amdgpu_ring *ring);
234 void (*emit_mem_sync)(struct amdgpu_ring *ring);
235 void (*emit_wave_limit)(struct amdgpu_ring *ring, bool enable);
236 void (*patch_cntl)(struct amdgpu_ring *ring, unsigned offset);
237 void (*patch_ce)(struct amdgpu_ring *ring, unsigned offset);
238 void (*patch_de)(struct amdgpu_ring *ring, unsigned offset);
239 };
240
241 struct amdgpu_ring {
242 struct amdgpu_device *adev;
243 const struct amdgpu_ring_funcs *funcs;
244 struct amdgpu_fence_driver fence_drv;
245 struct drm_gpu_scheduler sched;
246
247 struct amdgpu_bo *ring_obj;
248 volatile uint32_t *ring;
249 unsigned rptr_offs;
250 u64 rptr_gpu_addr;
251 volatile u32 *rptr_cpu_addr;
252 u64 wptr;
253 u64 wptr_old;
254 unsigned ring_size;
255 unsigned max_dw;
256 int count_dw;
257 uint64_t gpu_addr;
258 uint64_t ptr_mask;
259 uint32_t buf_mask;
260 u32 idx;
261 u32 xcc_id;
262 u32 xcp_id;
263 u32 me;
264 u32 pipe;
265 u32 queue;
266 struct amdgpu_bo *mqd_obj;
267 uint64_t mqd_gpu_addr;
268 void *mqd_ptr;
269 unsigned mqd_size;
270 uint64_t eop_gpu_addr;
271 u32 doorbell_index;
272 bool use_doorbell;
273 bool use_pollmem;
274 unsigned wptr_offs;
275 u64 wptr_gpu_addr;
276 volatile u32 *wptr_cpu_addr;
277 unsigned fence_offs;
278 u64 fence_gpu_addr;
279 volatile u32 *fence_cpu_addr;
280 uint64_t current_ctx;
281 char name[16];
282 u32 trail_seq;
283 unsigned trail_fence_offs;
284 u64 trail_fence_gpu_addr;
285 volatile u32 *trail_fence_cpu_addr;
286 unsigned cond_exe_offs;
287 u64 cond_exe_gpu_addr;
288 volatile u32 *cond_exe_cpu_addr;
289 unsigned vm_hub;
290 unsigned vm_inv_eng;
291 struct dma_fence *vmid_wait;
292 bool has_compute_vm_bug;
293 bool no_scheduler;
294 int hw_prio;
295 unsigned num_hw_submission;
296 atomic_t *sched_score;
297
298 /* used for mes */
299 bool is_mes_queue;
300 uint32_t hw_queue_id;
301 struct amdgpu_mes_ctx_data *mes_ctx;
302
303 bool is_sw_ring;
304 unsigned int entry_index;
305
306 };
307
308 #define amdgpu_ring_parse_cs(r, p, job, ib) ((r)->funcs->parse_cs((p), (job), (ib)))
309 #define amdgpu_ring_patch_cs_in_place(r, p, job, ib) ((r)->funcs->patch_cs_in_place((p), (job), (ib)))
310 #define amdgpu_ring_test_ring(r) (r)->funcs->test_ring((r))
311 #define amdgpu_ring_test_ib(r, t) ((r)->funcs->test_ib ? (r)->funcs->test_ib((r), (t)) : 0)
312 #define amdgpu_ring_get_rptr(r) (r)->funcs->get_rptr((r))
313 #define amdgpu_ring_get_wptr(r) (r)->funcs->get_wptr((r))
314 #define amdgpu_ring_set_wptr(r) (r)->funcs->set_wptr((r))
315 #define amdgpu_ring_emit_ib(r, job, ib, flags) ((r)->funcs->emit_ib((r), (job), (ib), (flags)))
316 #define amdgpu_ring_emit_pipeline_sync(r) (r)->funcs->emit_pipeline_sync((r))
317 #define amdgpu_ring_emit_vm_flush(r, vmid, addr) (r)->funcs->emit_vm_flush((r), (vmid), (addr))
318 #define amdgpu_ring_emit_fence(r, addr, seq, flags) (r)->funcs->emit_fence((r), (addr), (seq), (flags))
319 #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))
320 #define amdgpu_ring_emit_hdp_flush(r) (r)->funcs->emit_hdp_flush((r))
321 #define amdgpu_ring_emit_switch_buffer(r) (r)->funcs->emit_switch_buffer((r))
322 #define amdgpu_ring_emit_cntxcntl(r, d) (r)->funcs->emit_cntxcntl((r), (d))
323 #define amdgpu_ring_emit_gfx_shadow(r, s, c, g, i, v) ((r)->funcs->emit_gfx_shadow((r), (s), (c), (g), (i), (v)))
324 #define amdgpu_ring_emit_rreg(r, d, o) (r)->funcs->emit_rreg((r), (d), (o))
325 #define amdgpu_ring_emit_wreg(r, d, v) (r)->funcs->emit_wreg((r), (d), (v))
326 #define amdgpu_ring_emit_reg_wait(r, d, v, m) (r)->funcs->emit_reg_wait((r), (d), (v), (m))
327 #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))
328 #define amdgpu_ring_emit_frame_cntl(r, b, s) (r)->funcs->emit_frame_cntl((r), (b), (s))
329 #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
330 #define amdgpu_ring_init_cond_exec(r) (r)->funcs->init_cond_exec((r))
331 #define amdgpu_ring_patch_cond_exec(r,o) (r)->funcs->patch_cond_exec((r),(o))
332 #define amdgpu_ring_preempt_ib(r) (r)->funcs->preempt_ib(r)
333 #define amdgpu_ring_patch_cntl(r, o) ((r)->funcs->patch_cntl((r), (o)))
334 #define amdgpu_ring_patch_ce(r, o) ((r)->funcs->patch_ce((r), (o)))
335 #define amdgpu_ring_patch_de(r, o) ((r)->funcs->patch_de((r), (o)))
336
337 unsigned int amdgpu_ring_max_ibs(enum amdgpu_ring_type type);
338 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
339 void amdgpu_ring_ib_begin(struct amdgpu_ring *ring);
340 void amdgpu_ring_ib_end(struct amdgpu_ring *ring);
341 void amdgpu_ring_ib_on_emit_cntl(struct amdgpu_ring *ring);
342 void amdgpu_ring_ib_on_emit_ce(struct amdgpu_ring *ring);
343 void amdgpu_ring_ib_on_emit_de(struct amdgpu_ring *ring);
344
345 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count);
346 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib);
347 void amdgpu_ring_commit(struct amdgpu_ring *ring);
348 void amdgpu_ring_undo(struct amdgpu_ring *ring);
349 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
350 unsigned int max_dw, struct amdgpu_irq_src *irq_src,
351 unsigned int irq_type, unsigned int hw_prio,
352 atomic_t *sched_score);
353 void amdgpu_ring_fini(struct amdgpu_ring *ring);
354 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
355 uint32_t reg0, uint32_t val0,
356 uint32_t reg1, uint32_t val1);
357 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
358 struct dma_fence *fence);
359
amdgpu_ring_set_preempt_cond_exec(struct amdgpu_ring * ring,bool cond_exec)360 static inline void amdgpu_ring_set_preempt_cond_exec(struct amdgpu_ring *ring,
361 bool cond_exec)
362 {
363 *ring->cond_exe_cpu_addr = cond_exec;
364 }
365
amdgpu_ring_clear_ring(struct amdgpu_ring * ring)366 static inline void amdgpu_ring_clear_ring(struct amdgpu_ring *ring)
367 {
368 int i = 0;
369 while (i <= ring->buf_mask)
370 ring->ring[i++] = ring->funcs->nop;
371
372 }
373
amdgpu_ring_write(struct amdgpu_ring * ring,uint32_t v)374 static inline void amdgpu_ring_write(struct amdgpu_ring *ring, uint32_t v)
375 {
376 if (ring->count_dw <= 0)
377 DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
378 ring->ring[ring->wptr++ & ring->buf_mask] = v;
379 ring->wptr &= ring->ptr_mask;
380 ring->count_dw--;
381 }
382
amdgpu_ring_write_multiple(struct amdgpu_ring * ring,void * src,int count_dw)383 static inline void amdgpu_ring_write_multiple(struct amdgpu_ring *ring,
384 void *src, int count_dw)
385 {
386 unsigned occupied, chunk1, chunk2;
387 void *dst;
388
389 if (unlikely(ring->count_dw < count_dw))
390 DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
391
392 occupied = ring->wptr & ring->buf_mask;
393 dst = (void *)&ring->ring[occupied];
394 chunk1 = ring->buf_mask + 1 - occupied;
395 chunk1 = (chunk1 >= count_dw) ? count_dw : chunk1;
396 chunk2 = count_dw - chunk1;
397 chunk1 <<= 2;
398 chunk2 <<= 2;
399
400 if (chunk1)
401 memcpy(dst, src, chunk1);
402
403 if (chunk2) {
404 src += chunk1;
405 dst = (void *)ring->ring;
406 memcpy(dst, src, chunk2);
407 }
408
409 ring->wptr += count_dw;
410 ring->wptr &= ring->ptr_mask;
411 ring->count_dw -= count_dw;
412 }
413
414 #define amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset) \
415 (ring->is_mes_queue && ring->mes_ctx ? \
416 (ring->mes_ctx->meta_data_gpu_addr + offset) : 0)
417
418 #define amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset) \
419 (ring->is_mes_queue && ring->mes_ctx ? \
420 (void *)((uint8_t *)(ring->mes_ctx->meta_data_ptr) + offset) : \
421 NULL)
422
423 int amdgpu_ring_test_helper(struct amdgpu_ring *ring);
424
425 void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
426 struct amdgpu_ring *ring);
427
428 int amdgpu_ring_init_mqd(struct amdgpu_ring *ring);
429
amdgpu_ib_get_value(struct amdgpu_ib * ib,int idx)430 static inline u32 amdgpu_ib_get_value(struct amdgpu_ib *ib, int idx)
431 {
432 return ib->ptr[idx];
433 }
434
amdgpu_ib_set_value(struct amdgpu_ib * ib,int idx,uint32_t value)435 static inline void amdgpu_ib_set_value(struct amdgpu_ib *ib, int idx,
436 uint32_t value)
437 {
438 ib->ptr[idx] = value;
439 }
440
441 int amdgpu_ib_get(struct amdgpu_device *adev, struct amdgpu_vm *vm,
442 unsigned size,
443 enum amdgpu_ib_pool_type pool,
444 struct amdgpu_ib *ib);
445 void amdgpu_ib_free(struct amdgpu_device *adev, struct amdgpu_ib *ib,
446 struct dma_fence *f);
447 int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
448 struct amdgpu_ib *ibs, struct amdgpu_job *job,
449 struct dma_fence **f);
450 int amdgpu_ib_pool_init(struct amdgpu_device *adev);
451 void amdgpu_ib_pool_fini(struct amdgpu_device *adev);
452 int amdgpu_ib_ring_tests(struct amdgpu_device *adev);
453 bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring);
454 #endif
455