1 /*
2 * Copyright 2021 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 */
23
24 #include <linux/printk.h>
25 #include <linux/slab.h>
26 #include <linux/uaccess.h>
27 #include "kfd_priv.h"
28 #include "kfd_mqd_manager.h"
29 #include "v11_structs.h"
30 #include "gc/gc_11_0_0_offset.h"
31 #include "gc/gc_11_0_0_sh_mask.h"
32 #include "amdgpu_amdkfd.h"
33
get_mqd(void * mqd)34 static inline struct v11_compute_mqd *get_mqd(void *mqd)
35 {
36 return (struct v11_compute_mqd *)mqd;
37 }
38
get_sdma_mqd(void * mqd)39 static inline struct v11_sdma_mqd *get_sdma_mqd(void *mqd)
40 {
41 return (struct v11_sdma_mqd *)mqd;
42 }
43
update_cu_mask(struct mqd_manager * mm,void * mqd,struct mqd_update_info * minfo)44 static void update_cu_mask(struct mqd_manager *mm, void *mqd,
45 struct mqd_update_info *minfo)
46 {
47 struct v11_compute_mqd *m;
48 uint32_t se_mask[KFD_MAX_NUM_SE] = {0};
49 bool has_wa_flag = minfo && (minfo->update_flag & (UPDATE_FLAG_DBG_WA_ENABLE |
50 UPDATE_FLAG_DBG_WA_DISABLE));
51
52 if (!minfo || !(has_wa_flag || minfo->cu_mask.ptr))
53 return;
54
55 m = get_mqd(mqd);
56
57 if (has_wa_flag) {
58 uint32_t wa_mask =
59 (minfo->update_flag & UPDATE_FLAG_DBG_WA_ENABLE) ? 0xffff : 0xffffffff;
60
61 m->compute_static_thread_mgmt_se0 = wa_mask;
62 m->compute_static_thread_mgmt_se1 = wa_mask;
63 m->compute_static_thread_mgmt_se2 = wa_mask;
64 m->compute_static_thread_mgmt_se3 = wa_mask;
65 m->compute_static_thread_mgmt_se4 = wa_mask;
66 m->compute_static_thread_mgmt_se5 = wa_mask;
67 m->compute_static_thread_mgmt_se6 = wa_mask;
68 m->compute_static_thread_mgmt_se7 = wa_mask;
69
70 return;
71 }
72
73 mqd_symmetrically_map_cu_mask(mm,
74 minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask, 0);
75
76 m->compute_static_thread_mgmt_se0 = se_mask[0];
77 m->compute_static_thread_mgmt_se1 = se_mask[1];
78 m->compute_static_thread_mgmt_se2 = se_mask[2];
79 m->compute_static_thread_mgmt_se3 = se_mask[3];
80 m->compute_static_thread_mgmt_se4 = se_mask[4];
81 m->compute_static_thread_mgmt_se5 = se_mask[5];
82 m->compute_static_thread_mgmt_se6 = se_mask[6];
83 m->compute_static_thread_mgmt_se7 = se_mask[7];
84
85 pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n",
86 m->compute_static_thread_mgmt_se0,
87 m->compute_static_thread_mgmt_se1,
88 m->compute_static_thread_mgmt_se2,
89 m->compute_static_thread_mgmt_se3,
90 m->compute_static_thread_mgmt_se4,
91 m->compute_static_thread_mgmt_se5,
92 m->compute_static_thread_mgmt_se6,
93 m->compute_static_thread_mgmt_se7);
94 }
95
set_priority(struct v11_compute_mqd * m,struct queue_properties * q)96 static void set_priority(struct v11_compute_mqd *m, struct queue_properties *q)
97 {
98 m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
99 m->cp_hqd_queue_priority = q->priority;
100 }
101
allocate_mqd(struct kfd_node * node,struct queue_properties * q)102 static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
103 struct queue_properties *q)
104 {
105 struct kfd_mem_obj *mqd_mem_obj;
106 int size;
107
108 /*
109 * MES write to areas beyond MQD size. So allocate
110 * 1 PAGE_SIZE memory for MQD is MES is enabled.
111 */
112 if (node->kfd->shared_resources.enable_mes)
113 size = PAGE_SIZE;
114 else
115 size = sizeof(struct v11_compute_mqd);
116
117 if (kfd_gtt_sa_allocate(node, size, &mqd_mem_obj))
118 return NULL;
119
120 return mqd_mem_obj;
121 }
122
init_mqd(struct mqd_manager * mm,void ** mqd,struct kfd_mem_obj * mqd_mem_obj,uint64_t * gart_addr,struct queue_properties * q)123 static void init_mqd(struct mqd_manager *mm, void **mqd,
124 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
125 struct queue_properties *q)
126 {
127 uint64_t addr;
128 struct v11_compute_mqd *m;
129 int size;
130 uint32_t wa_mask = q->is_dbg_wa ? 0xffff : 0xffffffff;
131
132 m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr;
133 addr = mqd_mem_obj->gpu_addr;
134
135 if (mm->dev->kfd->shared_resources.enable_mes)
136 size = PAGE_SIZE;
137 else
138 size = sizeof(struct v11_compute_mqd);
139
140 memset(m, 0, size);
141
142 m->header = 0xC0310800;
143 m->compute_pipelinestat_enable = 1;
144
145 m->compute_static_thread_mgmt_se0 = wa_mask;
146 m->compute_static_thread_mgmt_se1 = wa_mask;
147 m->compute_static_thread_mgmt_se2 = wa_mask;
148 m->compute_static_thread_mgmt_se3 = wa_mask;
149 m->compute_static_thread_mgmt_se4 = wa_mask;
150 m->compute_static_thread_mgmt_se5 = wa_mask;
151 m->compute_static_thread_mgmt_se6 = wa_mask;
152 m->compute_static_thread_mgmt_se7 = wa_mask;
153
154 m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
155 0x55 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
156
157 m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
158 m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__UNORD_DISPATCH_MASK;
159 m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
160
161 m->cp_mqd_base_addr_lo = lower_32_bits(addr);
162 m->cp_mqd_base_addr_hi = upper_32_bits(addr);
163
164 m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
165 1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
166 1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
167
168 /* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the
169 * DISPATCH_PTR. This is required for the kfd debugger
170 */
171 m->cp_hqd_hq_status0 = 1 << 14;
172
173 /*
174 * GFX11 RS64 CPFW version >= 509 supports PCIe atomics support
175 * acknowledgment.
176 */
177 if (amdgpu_amdkfd_have_atomics_support(mm->dev->adev))
178 m->cp_hqd_hq_status0 |= 1 << 29;
179
180 if (q->format == KFD_QUEUE_FORMAT_AQL) {
181 m->cp_hqd_aql_control =
182 1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
183 }
184
185 if (mm->dev->kfd->cwsr_enabled) {
186 m->cp_hqd_persistent_state |=
187 (1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
188 m->cp_hqd_ctx_save_base_addr_lo =
189 lower_32_bits(q->ctx_save_restore_area_address);
190 m->cp_hqd_ctx_save_base_addr_hi =
191 upper_32_bits(q->ctx_save_restore_area_address);
192 m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
193 m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
194 m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
195 m->cp_hqd_wg_state_offset = q->ctl_stack_size;
196 }
197
198 *mqd = m;
199 if (gart_addr)
200 *gart_addr = addr;
201 mm->update_mqd(mm, m, q, NULL);
202 }
203
load_mqd(struct mqd_manager * mm,void * mqd,uint32_t pipe_id,uint32_t queue_id,struct queue_properties * p,struct mm_struct * mms)204 static int load_mqd(struct mqd_manager *mm, void *mqd,
205 uint32_t pipe_id, uint32_t queue_id,
206 struct queue_properties *p, struct mm_struct *mms)
207 {
208 int r = 0;
209 /* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
210 uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
211
212 r = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id,
213 (uint32_t __user *)p->write_ptr,
214 wptr_shift, 0, mms, 0);
215 return r;
216 }
217
update_mqd(struct mqd_manager * mm,void * mqd,struct queue_properties * q,struct mqd_update_info * minfo)218 static void update_mqd(struct mqd_manager *mm, void *mqd,
219 struct queue_properties *q,
220 struct mqd_update_info *minfo)
221 {
222 struct v11_compute_mqd *m;
223
224 m = get_mqd(mqd);
225
226 m->cp_hqd_pq_control &= ~CP_HQD_PQ_CONTROL__QUEUE_SIZE_MASK;
227 m->cp_hqd_pq_control |=
228 ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1;
229 pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
230
231 m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
232 m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
233
234 m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
235 m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
236 m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
237 m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
238
239 m->cp_hqd_pq_doorbell_control =
240 q->doorbell_off <<
241 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
242 pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
243 m->cp_hqd_pq_doorbell_control);
244
245 m->cp_hqd_ib_control = 3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT;
246
247 /*
248 * HW does not clamp this field correctly. Maximum EOP queue size
249 * is constrained by per-SE EOP done signal count, which is 8-bit.
250 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
251 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
252 * is safe, giving a maximum field value of 0xA.
253 */
254 m->cp_hqd_eop_control = min(0xA,
255 ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1);
256 m->cp_hqd_eop_base_addr_lo =
257 lower_32_bits(q->eop_ring_buffer_address >> 8);
258 m->cp_hqd_eop_base_addr_hi =
259 upper_32_bits(q->eop_ring_buffer_address >> 8);
260
261 m->cp_hqd_iq_timer = 0;
262
263 m->cp_hqd_vmid = q->vmid;
264
265 if (q->format == KFD_QUEUE_FORMAT_AQL) {
266 /* GC 10 removed WPP_CLAMP from PQ Control */
267 m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
268 2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
269 1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT ;
270 m->cp_hqd_pq_doorbell_control |=
271 1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
272 }
273 if (mm->dev->kfd->cwsr_enabled)
274 m->cp_hqd_ctx_save_control = 0;
275
276 update_cu_mask(mm, mqd, minfo);
277 set_priority(m, q);
278
279 q->is_active = QUEUE_IS_ACTIVE(*q);
280 }
281
check_preemption_failed(struct mqd_manager * mm,void * mqd)282 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
283 {
284 struct v11_compute_mqd *m = (struct v11_compute_mqd *)mqd;
285
286 return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
287 }
288
get_wave_state(struct mqd_manager * mm,void * mqd,struct queue_properties * q,void __user * ctl_stack,u32 * ctl_stack_used_size,u32 * save_area_used_size)289 static int get_wave_state(struct mqd_manager *mm, void *mqd,
290 struct queue_properties *q,
291 void __user *ctl_stack,
292 u32 *ctl_stack_used_size,
293 u32 *save_area_used_size)
294 {
295 struct v11_compute_mqd *m;
296 struct kfd_context_save_area_header header;
297
298 m = get_mqd(mqd);
299
300 /* Control stack is written backwards, while workgroup context data
301 * is written forwards. Both starts from m->cp_hqd_cntl_stack_size.
302 * Current position is at m->cp_hqd_cntl_stack_offset and
303 * m->cp_hqd_wg_state_offset, respectively.
304 */
305 *ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
306 m->cp_hqd_cntl_stack_offset;
307 *save_area_used_size = m->cp_hqd_wg_state_offset -
308 m->cp_hqd_cntl_stack_size;
309
310 /* Control stack is not copied to user mode for GFXv11 because
311 * it's part of the context save area that is already
312 * accessible to user mode
313 */
314 header.wave_state.control_stack_size = *ctl_stack_used_size;
315 header.wave_state.wave_state_size = *save_area_used_size;
316
317 header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset;
318 header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset;
319
320 if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state)))
321 return -EFAULT;
322
323 return 0;
324 }
325
checkpoint_mqd(struct mqd_manager * mm,void * mqd,void * mqd_dst,void * ctl_stack_dst)326 static void checkpoint_mqd(struct mqd_manager *mm, void *mqd, void *mqd_dst, void *ctl_stack_dst)
327 {
328 struct v11_compute_mqd *m;
329
330 m = get_mqd(mqd);
331
332 memcpy(mqd_dst, m, sizeof(struct v11_compute_mqd));
333 }
334
restore_mqd(struct mqd_manager * mm,void ** mqd,struct kfd_mem_obj * mqd_mem_obj,uint64_t * gart_addr,struct queue_properties * qp,const void * mqd_src,const void * ctl_stack_src,const u32 ctl_stack_size)335 static void restore_mqd(struct mqd_manager *mm, void **mqd,
336 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
337 struct queue_properties *qp,
338 const void *mqd_src,
339 const void *ctl_stack_src, const u32 ctl_stack_size)
340 {
341 uint64_t addr;
342 struct v11_compute_mqd *m;
343
344 m = (struct v11_compute_mqd *) mqd_mem_obj->cpu_ptr;
345 addr = mqd_mem_obj->gpu_addr;
346
347 memcpy(m, mqd_src, sizeof(*m));
348
349 *mqd = m;
350 if (gart_addr)
351 *gart_addr = addr;
352
353 m->cp_hqd_pq_doorbell_control =
354 qp->doorbell_off <<
355 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
356 pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
357 m->cp_hqd_pq_doorbell_control);
358
359 qp->is_active = 0;
360 }
361
362
init_mqd_hiq(struct mqd_manager * mm,void ** mqd,struct kfd_mem_obj * mqd_mem_obj,uint64_t * gart_addr,struct queue_properties * q)363 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
364 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
365 struct queue_properties *q)
366 {
367 struct v11_compute_mqd *m;
368
369 init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
370
371 m = get_mqd(*mqd);
372
373 m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
374 1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
375 }
376
destroy_hiq_mqd(struct mqd_manager * mm,void * mqd,enum kfd_preempt_type type,unsigned int timeout,uint32_t pipe_id,uint32_t queue_id)377 static int destroy_hiq_mqd(struct mqd_manager *mm, void *mqd,
378 enum kfd_preempt_type type, unsigned int timeout,
379 uint32_t pipe_id, uint32_t queue_id)
380 {
381 int err;
382 struct v11_compute_mqd *m;
383 u32 doorbell_off;
384
385 m = get_mqd(mqd);
386
387 doorbell_off = m->cp_hqd_pq_doorbell_control >>
388 CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
389
390 err = amdgpu_amdkfd_unmap_hiq(mm->dev->adev, doorbell_off, 0);
391 if (err)
392 pr_debug("Destroy HIQ MQD failed: %d\n", err);
393
394 return err;
395 }
396
init_mqd_sdma(struct mqd_manager * mm,void ** mqd,struct kfd_mem_obj * mqd_mem_obj,uint64_t * gart_addr,struct queue_properties * q)397 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
398 struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
399 struct queue_properties *q)
400 {
401 struct v11_sdma_mqd *m;
402 int size;
403
404 m = (struct v11_sdma_mqd *) mqd_mem_obj->cpu_ptr;
405
406 if (mm->dev->kfd->shared_resources.enable_mes)
407 size = PAGE_SIZE;
408 else
409 size = sizeof(struct v11_sdma_mqd);
410
411 memset(m, 0, size);
412 *mqd = m;
413 if (gart_addr)
414 *gart_addr = mqd_mem_obj->gpu_addr;
415
416 mm->update_mqd(mm, m, q, NULL);
417 }
418
419 #define SDMA_RLC_DUMMY_DEFAULT 0xf
420
update_mqd_sdma(struct mqd_manager * mm,void * mqd,struct queue_properties * q,struct mqd_update_info * minfo)421 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
422 struct queue_properties *q,
423 struct mqd_update_info *minfo)
424 {
425 struct v11_sdma_mqd *m;
426
427 m = get_sdma_mqd(mqd);
428 m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1)
429 << SDMA0_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
430 q->vmid << SDMA0_QUEUE0_RB_CNTL__RB_VMID__SHIFT |
431 1 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
432 6 << SDMA0_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
433 1 << SDMA0_QUEUE0_RB_CNTL__F32_WPTR_POLL_ENABLE__SHIFT;
434
435 m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
436 m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
437 m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
438 m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
439 m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
440 m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
441 m->sdmax_rlcx_doorbell_offset =
442 q->doorbell_off << SDMA0_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
443
444 m->sdmax_rlcx_sched_cntl = (amdgpu_sdma_phase_quantum
445 << SDMA0_QUEUE0_SCHEDULE_CNTL__CONTEXT_QUANTUM__SHIFT)
446 & SDMA0_QUEUE0_SCHEDULE_CNTL__CONTEXT_QUANTUM_MASK;
447
448 m->sdma_engine_id = q->sdma_engine_id;
449 m->sdma_queue_id = q->sdma_queue_id;
450 m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
451
452 q->is_active = QUEUE_IS_ACTIVE(*q);
453 }
454
455 #if defined(CONFIG_DEBUG_FS)
456
debugfs_show_mqd(struct seq_file * m,void * data)457 static int debugfs_show_mqd(struct seq_file *m, void *data)
458 {
459 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
460 data, sizeof(struct v11_compute_mqd), false);
461 return 0;
462 }
463
debugfs_show_mqd_sdma(struct seq_file * m,void * data)464 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
465 {
466 seq_hex_dump(m, " ", DUMP_PREFIX_OFFSET, 32, 4,
467 data, sizeof(struct v11_sdma_mqd), false);
468 return 0;
469 }
470
471 #endif
472
mqd_manager_init_v11(enum KFD_MQD_TYPE type,struct kfd_node * dev)473 struct mqd_manager *mqd_manager_init_v11(enum KFD_MQD_TYPE type,
474 struct kfd_node *dev)
475 {
476 struct mqd_manager *mqd;
477
478 if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
479 return NULL;
480
481 mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
482 if (!mqd)
483 return NULL;
484
485 mqd->dev = dev;
486
487 switch (type) {
488 case KFD_MQD_TYPE_CP:
489 pr_debug("%s@%i\n", __func__, __LINE__);
490 mqd->allocate_mqd = allocate_mqd;
491 mqd->init_mqd = init_mqd;
492 mqd->free_mqd = kfd_free_mqd_cp;
493 mqd->load_mqd = load_mqd;
494 mqd->update_mqd = update_mqd;
495 mqd->destroy_mqd = kfd_destroy_mqd_cp;
496 mqd->is_occupied = kfd_is_occupied_cp;
497 mqd->mqd_size = sizeof(struct v11_compute_mqd);
498 mqd->get_wave_state = get_wave_state;
499 mqd->mqd_stride = kfd_mqd_stride;
500 mqd->checkpoint_mqd = checkpoint_mqd;
501 mqd->restore_mqd = restore_mqd;
502 #if defined(CONFIG_DEBUG_FS)
503 mqd->debugfs_show_mqd = debugfs_show_mqd;
504 #endif
505 pr_debug("%s@%i\n", __func__, __LINE__);
506 break;
507 case KFD_MQD_TYPE_HIQ:
508 pr_debug("%s@%i\n", __func__, __LINE__);
509 mqd->allocate_mqd = allocate_hiq_mqd;
510 mqd->init_mqd = init_mqd_hiq;
511 mqd->free_mqd = free_mqd_hiq_sdma;
512 mqd->load_mqd = kfd_hiq_load_mqd_kiq;
513 mqd->update_mqd = update_mqd;
514 mqd->destroy_mqd = destroy_hiq_mqd;
515 mqd->is_occupied = kfd_is_occupied_cp;
516 mqd->mqd_size = sizeof(struct v11_compute_mqd);
517 mqd->mqd_stride = kfd_mqd_stride;
518 #if defined(CONFIG_DEBUG_FS)
519 mqd->debugfs_show_mqd = debugfs_show_mqd;
520 #endif
521 mqd->check_preemption_failed = check_preemption_failed;
522 pr_debug("%s@%i\n", __func__, __LINE__);
523 break;
524 case KFD_MQD_TYPE_DIQ:
525 mqd->allocate_mqd = allocate_mqd;
526 mqd->init_mqd = init_mqd_hiq;
527 mqd->free_mqd = kfd_free_mqd_cp;
528 mqd->load_mqd = load_mqd;
529 mqd->update_mqd = update_mqd;
530 mqd->destroy_mqd = kfd_destroy_mqd_cp;
531 mqd->is_occupied = kfd_is_occupied_cp;
532 mqd->mqd_size = sizeof(struct v11_compute_mqd);
533 #if defined(CONFIG_DEBUG_FS)
534 mqd->debugfs_show_mqd = debugfs_show_mqd;
535 #endif
536 break;
537 case KFD_MQD_TYPE_SDMA:
538 pr_debug("%s@%i\n", __func__, __LINE__);
539 mqd->allocate_mqd = allocate_sdma_mqd;
540 mqd->init_mqd = init_mqd_sdma;
541 mqd->free_mqd = free_mqd_hiq_sdma;
542 mqd->load_mqd = kfd_load_mqd_sdma;
543 mqd->update_mqd = update_mqd_sdma;
544 mqd->destroy_mqd = kfd_destroy_mqd_sdma;
545 mqd->is_occupied = kfd_is_occupied_sdma;
546 mqd->checkpoint_mqd = checkpoint_mqd;
547 mqd->restore_mqd = restore_mqd;
548 mqd->mqd_size = sizeof(struct v11_sdma_mqd);
549 mqd->mqd_stride = kfd_mqd_stride;
550 #if defined(CONFIG_DEBUG_FS)
551 mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
552 #endif
553 /*
554 * To allocate SDMA MQDs by generic functions
555 * when MES is enabled.
556 */
557 if (dev->kfd->shared_resources.enable_mes) {
558 mqd->allocate_mqd = allocate_mqd;
559 mqd->free_mqd = kfd_free_mqd_cp;
560 }
561 pr_debug("%s@%i\n", __func__, __LINE__);
562 break;
563 default:
564 kfree(mqd);
565 return NULL;
566 }
567
568 return mqd;
569 }
570