xref: /openbmc/linux/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c (revision 1d8355ad922423c9f765a644ed04526a6273d9ee)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2016-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  *
23  */
24 
25 #include <linux/printk.h>
26 #include <linux/slab.h>
27 #include <linux/uaccess.h>
28 #include "kfd_priv.h"
29 #include "kfd_mqd_manager.h"
30 #include "v9_structs.h"
31 #include "gc/gc_9_0_offset.h"
32 #include "gc/gc_9_0_sh_mask.h"
33 #include "sdma0/sdma0_4_0_sh_mask.h"
34 #include "amdgpu_amdkfd.h"
35 
36 static inline struct v9_mqd *get_mqd(void *mqd)
37 {
38 	return (struct v9_mqd *)mqd;
39 }
40 
41 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
42 {
43 	return (struct v9_sdma_mqd *)mqd;
44 }
45 
46 static void update_cu_mask(struct mqd_manager *mm, void *mqd,
47 			struct mqd_update_info *minfo)
48 {
49 	struct v9_mqd *m;
50 	uint32_t se_mask[KFD_MAX_NUM_SE] = {0};
51 
52 	if (!minfo || (minfo->update_flag != UPDATE_FLAG_CU_MASK) ||
53 	    !minfo->cu_mask.ptr)
54 		return;
55 
56 	mqd_symmetrically_map_cu_mask(mm,
57 		minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask);
58 
59 	m = get_mqd(mqd);
60 	m->compute_static_thread_mgmt_se0 = se_mask[0];
61 	m->compute_static_thread_mgmt_se1 = se_mask[1];
62 	m->compute_static_thread_mgmt_se2 = se_mask[2];
63 	m->compute_static_thread_mgmt_se3 = se_mask[3];
64 	m->compute_static_thread_mgmt_se4 = se_mask[4];
65 	m->compute_static_thread_mgmt_se5 = se_mask[5];
66 	m->compute_static_thread_mgmt_se6 = se_mask[6];
67 	m->compute_static_thread_mgmt_se7 = se_mask[7];
68 
69 	pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n",
70 		m->compute_static_thread_mgmt_se0,
71 		m->compute_static_thread_mgmt_se1,
72 		m->compute_static_thread_mgmt_se2,
73 		m->compute_static_thread_mgmt_se3,
74 		m->compute_static_thread_mgmt_se4,
75 		m->compute_static_thread_mgmt_se5,
76 		m->compute_static_thread_mgmt_se6,
77 		m->compute_static_thread_mgmt_se7);
78 }
79 
80 static void set_priority(struct v9_mqd *m, struct queue_properties *q)
81 {
82 	m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
83 	m->cp_hqd_queue_priority = q->priority;
84 }
85 
86 static struct kfd_mem_obj *allocate_mqd(struct kfd_dev *kfd,
87 		struct queue_properties *q)
88 {
89 	int retval;
90 	struct kfd_mem_obj *mqd_mem_obj = NULL;
91 
92 	/* For V9 only, due to a HW bug, the control stack of a user mode
93 	 * compute queue needs to be allocated just behind the page boundary
94 	 * of its regular MQD buffer. So we allocate an enlarged MQD buffer:
95 	 * the first page of the buffer serves as the regular MQD buffer
96 	 * purpose and the remaining is for control stack. Although the two
97 	 * parts are in the same buffer object, they need different memory
98 	 * types: MQD part needs UC (uncached) as usual, while control stack
99 	 * needs NC (non coherent), which is different from the UC type which
100 	 * is used when control stack is allocated in user space.
101 	 *
102 	 * Because of all those, we use the gtt allocation function instead
103 	 * of sub-allocation function for this enlarged MQD buffer. Moreover,
104 	 * in order to achieve two memory types in a single buffer object, we
105 	 * pass a special bo flag AMDGPU_GEM_CREATE_CP_MQD_GFX9 to instruct
106 	 * amdgpu memory functions to do so.
107 	 */
108 	if (kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) {
109 		mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
110 		if (!mqd_mem_obj)
111 			return NULL;
112 		retval = amdgpu_amdkfd_alloc_gtt_mem(kfd->adev,
113 			ALIGN(q->ctl_stack_size, PAGE_SIZE) +
114 				ALIGN(sizeof(struct v9_mqd), PAGE_SIZE),
115 			&(mqd_mem_obj->gtt_mem),
116 			&(mqd_mem_obj->gpu_addr),
117 			(void *)&(mqd_mem_obj->cpu_ptr), true);
118 
119 		if (retval) {
120 			kfree(mqd_mem_obj);
121 			return NULL;
122 		}
123 	} else {
124 		retval = kfd_gtt_sa_allocate(kfd, sizeof(struct v9_mqd),
125 				&mqd_mem_obj);
126 		if (retval)
127 			return NULL;
128 	}
129 
130 	return mqd_mem_obj;
131 }
132 
133 static void init_mqd(struct mqd_manager *mm, void **mqd,
134 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
135 			struct queue_properties *q)
136 {
137 	uint64_t addr;
138 	struct v9_mqd *m;
139 	struct amdgpu_device *adev = (struct amdgpu_device *)mm->dev->adev;
140 
141 	m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr;
142 	addr = mqd_mem_obj->gpu_addr;
143 
144 	memset(m, 0, sizeof(struct v9_mqd));
145 
146 	m->header = 0xC0310800;
147 	m->compute_pipelinestat_enable = 1;
148 	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
149 	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
150 	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
151 	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
152 	m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF;
153 	m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF;
154 	m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF;
155 	m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF;
156 
157 	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
158 			0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
159 
160 	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
161 
162 	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
163 	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
164 
165 	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
166 			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
167 			1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
168 
169 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
170 		m->cp_hqd_aql_control =
171 			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
172 		if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) {
173 			/* On GC 9.4.3, DW 41 is re-purposed as
174 			 * compute_tg_chunk_size.
175 			 * TODO: review this setting when active CUs in the
176 			 * partition play a role
177 			 */
178 			m->compute_static_thread_mgmt_se6 = 1;
179 		}
180 	} else {
181 		/* PM4 queue */
182 		if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 3)) {
183 			m->compute_static_thread_mgmt_se6 = 0;
184 			/* TODO: program pm4_target_xcc */
185 		}
186 	}
187 
188 	if (q->tba_addr) {
189 		m->compute_pgm_rsrc2 |=
190 			(1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT);
191 	}
192 
193 	if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address) {
194 		m->cp_hqd_persistent_state |=
195 			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
196 		m->cp_hqd_ctx_save_base_addr_lo =
197 			lower_32_bits(q->ctx_save_restore_area_address);
198 		m->cp_hqd_ctx_save_base_addr_hi =
199 			upper_32_bits(q->ctx_save_restore_area_address);
200 		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
201 		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
202 		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
203 		m->cp_hqd_wg_state_offset = q->ctl_stack_size;
204 	}
205 
206 	*mqd = m;
207 	if (gart_addr)
208 		*gart_addr = addr;
209 	mm->update_mqd(mm, m, q, NULL);
210 }
211 
212 static int load_mqd(struct mqd_manager *mm, void *mqd,
213 			uint32_t pipe_id, uint32_t queue_id,
214 			struct queue_properties *p, struct mm_struct *mms)
215 {
216 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
217 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
218 
219 	return mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id,
220 					  (uint32_t __user *)p->write_ptr,
221 					  wptr_shift, 0, mms);
222 }
223 
224 static void update_mqd(struct mqd_manager *mm, void *mqd,
225 			struct queue_properties *q,
226 			struct mqd_update_info *minfo)
227 {
228 	struct amdgpu_device *adev = (struct amdgpu_device *)mm->dev->adev;
229 	struct v9_mqd *m;
230 
231 	m = get_mqd(mqd);
232 
233 	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
234 	m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1;
235 	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
236 
237 	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
238 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
239 
240 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
241 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
242 	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
243 	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
244 
245 	m->cp_hqd_pq_doorbell_control =
246 		q->doorbell_off <<
247 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
248 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
249 			m->cp_hqd_pq_doorbell_control);
250 
251 	m->cp_hqd_ib_control =
252 		3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT |
253 		1 << CP_HQD_IB_CONTROL__IB_EXE_DISABLE__SHIFT;
254 
255 	/*
256 	 * HW does not clamp this field correctly. Maximum EOP queue size
257 	 * is constrained by per-SE EOP done signal count, which is 8-bit.
258 	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
259 	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
260 	 * is safe, giving a maximum field value of 0xA.
261 	 */
262 	m->cp_hqd_eop_control = min(0xA,
263 		order_base_2(q->eop_ring_buffer_size / 4) - 1);
264 	m->cp_hqd_eop_base_addr_lo =
265 			lower_32_bits(q->eop_ring_buffer_address >> 8);
266 	m->cp_hqd_eop_base_addr_hi =
267 			upper_32_bits(q->eop_ring_buffer_address >> 8);
268 
269 	m->cp_hqd_iq_timer = 0;
270 
271 	m->cp_hqd_vmid = q->vmid;
272 
273 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
274 		m->cp_hqd_pq_control |=
275 				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
276 				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT |
277 				1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT;
278 		if (adev->ip_versions[GC_HWIP][0] != IP_VERSION(9, 4, 3))
279 			m->cp_hqd_pq_control |=
280 				 CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
281 		m->cp_hqd_pq_doorbell_control |= 1 <<
282 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
283 	}
284 	if (mm->dev->cwsr_enabled && q->ctx_save_restore_area_address)
285 		m->cp_hqd_ctx_save_control = 0;
286 
287 	update_cu_mask(mm, mqd, minfo);
288 	set_priority(m, q);
289 
290 	q->is_active = QUEUE_IS_ACTIVE(*q);
291 }
292 
293 
294 static uint32_t read_doorbell_id(void *mqd)
295 {
296 	struct v9_mqd *m = (struct v9_mqd *)mqd;
297 
298 	return m->queue_doorbell_id0;
299 }
300 
301 static int get_wave_state(struct mqd_manager *mm, void *mqd,
302 			  void __user *ctl_stack,
303 			  u32 *ctl_stack_used_size,
304 			  u32 *save_area_used_size)
305 {
306 	struct v9_mqd *m;
307 
308 	/* Control stack is located one page after MQD. */
309 	void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE);
310 
311 	m = get_mqd(mqd);
312 
313 	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
314 		m->cp_hqd_cntl_stack_offset;
315 	*save_area_used_size = m->cp_hqd_wg_state_offset -
316 		m->cp_hqd_cntl_stack_size;
317 
318 	if (copy_to_user(ctl_stack, mqd_ctl_stack, m->cp_hqd_cntl_stack_size))
319 		return -EFAULT;
320 
321 	return 0;
322 }
323 
324 static void get_checkpoint_info(struct mqd_manager *mm, void *mqd, u32 *ctl_stack_size)
325 {
326 	struct v9_mqd *m = get_mqd(mqd);
327 
328 	*ctl_stack_size = m->cp_hqd_cntl_stack_size;
329 }
330 
331 static void checkpoint_mqd(struct mqd_manager *mm, void *mqd, void *mqd_dst, void *ctl_stack_dst)
332 {
333 	struct v9_mqd *m;
334 	/* Control stack is located one page after MQD. */
335 	void *ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE);
336 
337 	m = get_mqd(mqd);
338 
339 	memcpy(mqd_dst, m, sizeof(struct v9_mqd));
340 	memcpy(ctl_stack_dst, ctl_stack, m->cp_hqd_cntl_stack_size);
341 }
342 
343 static void restore_mqd(struct mqd_manager *mm, void **mqd,
344 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
345 			struct queue_properties *qp,
346 			const void *mqd_src,
347 			const void *ctl_stack_src, u32 ctl_stack_size)
348 {
349 	uint64_t addr;
350 	struct v9_mqd *m;
351 	void *ctl_stack;
352 
353 	m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr;
354 	addr = mqd_mem_obj->gpu_addr;
355 
356 	memcpy(m, mqd_src, sizeof(*m));
357 
358 	*mqd = m;
359 	if (gart_addr)
360 		*gart_addr = addr;
361 
362 	/* Control stack is located one page after MQD. */
363 	ctl_stack = (void *)((uintptr_t)*mqd + PAGE_SIZE);
364 	memcpy(ctl_stack, ctl_stack_src, ctl_stack_size);
365 
366 	m->cp_hqd_pq_doorbell_control =
367 		qp->doorbell_off <<
368 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
369 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
370 				m->cp_hqd_pq_doorbell_control);
371 
372 	qp->is_active = 0;
373 }
374 
375 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
376 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
377 			struct queue_properties *q)
378 {
379 	struct v9_mqd *m;
380 
381 	init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
382 
383 	m = get_mqd(*mqd);
384 
385 	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
386 			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
387 }
388 
389 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
390 		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
391 		struct queue_properties *q)
392 {
393 	struct v9_sdma_mqd *m;
394 
395 	m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr;
396 
397 	memset(m, 0, sizeof(struct v9_sdma_mqd));
398 
399 	*mqd = m;
400 	if (gart_addr)
401 		*gart_addr = mqd_mem_obj->gpu_addr;
402 
403 	mm->update_mqd(mm, m, q, NULL);
404 }
405 
406 #define SDMA_RLC_DUMMY_DEFAULT 0xf
407 
408 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
409 			struct queue_properties *q,
410 			struct mqd_update_info *minfo)
411 {
412 	struct v9_sdma_mqd *m;
413 
414 	m = get_sdma_mqd(mqd);
415 	m->sdmax_rlcx_rb_cntl = order_base_2(q->queue_size / 4)
416 		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
417 		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
418 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
419 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
420 
421 	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
422 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
423 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
424 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
425 	m->sdmax_rlcx_doorbell_offset =
426 		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
427 
428 	m->sdma_engine_id = q->sdma_engine_id;
429 	m->sdma_queue_id = q->sdma_queue_id;
430 	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
431 
432 	q->is_active = QUEUE_IS_ACTIVE(*q);
433 }
434 
435 static void checkpoint_mqd_sdma(struct mqd_manager *mm,
436 				void *mqd,
437 				void *mqd_dst,
438 				void *ctl_stack_dst)
439 {
440 	struct v9_sdma_mqd *m;
441 
442 	m = get_sdma_mqd(mqd);
443 
444 	memcpy(mqd_dst, m, sizeof(struct v9_sdma_mqd));
445 }
446 
447 static void restore_mqd_sdma(struct mqd_manager *mm, void **mqd,
448 			     struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
449 			     struct queue_properties *qp,
450 			     const void *mqd_src,
451 			     const void *ctl_stack_src, const u32 ctl_stack_size)
452 {
453 	uint64_t addr;
454 	struct v9_sdma_mqd *m;
455 
456 	m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr;
457 	addr = mqd_mem_obj->gpu_addr;
458 
459 	memcpy(m, mqd_src, sizeof(*m));
460 
461 	m->sdmax_rlcx_doorbell_offset =
462 		qp->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
463 
464 	*mqd = m;
465 	if (gart_addr)
466 		*gart_addr = addr;
467 
468 	qp->is_active = 0;
469 }
470 
471 #if defined(CONFIG_DEBUG_FS)
472 
473 static int debugfs_show_mqd(struct seq_file *m, void *data)
474 {
475 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
476 		     data, sizeof(struct v9_mqd), false);
477 	return 0;
478 }
479 
480 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
481 {
482 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
483 		     data, sizeof(struct v9_sdma_mqd), false);
484 	return 0;
485 }
486 
487 #endif
488 
489 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
490 		struct kfd_dev *dev)
491 {
492 	struct mqd_manager *mqd;
493 
494 	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
495 		return NULL;
496 
497 	mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
498 	if (!mqd)
499 		return NULL;
500 
501 	mqd->dev = dev;
502 
503 	switch (type) {
504 	case KFD_MQD_TYPE_CP:
505 		mqd->allocate_mqd = allocate_mqd;
506 		mqd->init_mqd = init_mqd;
507 		mqd->free_mqd = kfd_free_mqd_cp;
508 		mqd->load_mqd = load_mqd;
509 		mqd->update_mqd = update_mqd;
510 		mqd->destroy_mqd = kfd_destroy_mqd_cp;
511 		mqd->is_occupied = kfd_is_occupied_cp;
512 		mqd->get_wave_state = get_wave_state;
513 		mqd->get_checkpoint_info = get_checkpoint_info;
514 		mqd->checkpoint_mqd = checkpoint_mqd;
515 		mqd->restore_mqd = restore_mqd;
516 		mqd->mqd_size = sizeof(struct v9_mqd);
517 #if defined(CONFIG_DEBUG_FS)
518 		mqd->debugfs_show_mqd = debugfs_show_mqd;
519 #endif
520 		break;
521 	case KFD_MQD_TYPE_HIQ:
522 		mqd->allocate_mqd = allocate_hiq_mqd;
523 		mqd->init_mqd = init_mqd_hiq;
524 		mqd->free_mqd = free_mqd_hiq_sdma;
525 		mqd->load_mqd = kfd_hiq_load_mqd_kiq;
526 		mqd->update_mqd = update_mqd;
527 		mqd->destroy_mqd = kfd_destroy_mqd_cp;
528 		mqd->is_occupied = kfd_is_occupied_cp;
529 		mqd->mqd_size = sizeof(struct v9_mqd);
530 #if defined(CONFIG_DEBUG_FS)
531 		mqd->debugfs_show_mqd = debugfs_show_mqd;
532 #endif
533 		mqd->read_doorbell_id = read_doorbell_id;
534 		break;
535 	case KFD_MQD_TYPE_DIQ:
536 		mqd->allocate_mqd = allocate_mqd;
537 		mqd->init_mqd = init_mqd_hiq;
538 		mqd->free_mqd = kfd_free_mqd_cp;
539 		mqd->load_mqd = load_mqd;
540 		mqd->update_mqd = update_mqd;
541 		mqd->destroy_mqd = kfd_destroy_mqd_cp;
542 		mqd->is_occupied = kfd_is_occupied_cp;
543 		mqd->mqd_size = sizeof(struct v9_mqd);
544 #if defined(CONFIG_DEBUG_FS)
545 		mqd->debugfs_show_mqd = debugfs_show_mqd;
546 #endif
547 		break;
548 	case KFD_MQD_TYPE_SDMA:
549 		mqd->allocate_mqd = allocate_sdma_mqd;
550 		mqd->init_mqd = init_mqd_sdma;
551 		mqd->free_mqd = free_mqd_hiq_sdma;
552 		mqd->load_mqd = kfd_load_mqd_sdma;
553 		mqd->update_mqd = update_mqd_sdma;
554 		mqd->destroy_mqd = kfd_destroy_mqd_sdma;
555 		mqd->is_occupied = kfd_is_occupied_sdma;
556 		mqd->checkpoint_mqd = checkpoint_mqd_sdma;
557 		mqd->restore_mqd = restore_mqd_sdma;
558 		mqd->mqd_size = sizeof(struct v9_sdma_mqd);
559 #if defined(CONFIG_DEBUG_FS)
560 		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
561 #endif
562 		break;
563 	default:
564 		kfree(mqd);
565 		return NULL;
566 	}
567 
568 	return mqd;
569 }
570