xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v12_1.c (revision bf4afc53b77aeaa48b5409da5c8da6bb4eff7f43)
1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2025 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 "v12_structs.h"
31 #include "gc/gc_12_1_0_sh_mask.h"
32 #include "amdgpu_amdkfd.h"
33 #include "kfd_device_queue_manager.h"
34 
35 static inline struct v12_1_compute_mqd *get_mqd(void *mqd)
36 {
37 	return (struct v12_1_compute_mqd *)mqd;
38 }
39 
40 static inline struct v12_sdma_mqd *get_sdma_mqd(void *mqd)
41 {
42 	return (struct v12_sdma_mqd *)mqd;
43 }
44 
45 static void mqd_symmetrically_map_cu_mask_v12_1(struct mqd_manager *mm,
46 		const uint32_t *cu_mask, uint32_t cu_mask_count,
47 		uint32_t *se_mask, uint32_t inst)
48 {
49 	struct amdgpu_cu_info *cu_info = &mm->dev->adev->gfx.cu_info;
50 	struct amdgpu_gfx_config *gfx_info = &mm->dev->adev->gfx.config;
51 	uint32_t cu_per_sh[2][2] = {0};
52 	uint32_t en_mask = 0x3;
53 	int i, se, sh, cu, cu_inc = 0;
54 	uint32_t cu_active_per_node;
55 	int inc = NUM_XCC(mm->dev->xcc_mask);
56 	int xcc_inst = inst + ffs(mm->dev->xcc_mask) - 1;
57 
58 	cu_active_per_node = cu_info->number / mm->dev->kfd->num_nodes;
59 	if (cu_mask_count > cu_active_per_node)
60 		cu_mask_count = cu_active_per_node;
61 
62 	/*
63 	 * Count active CUs per SE/SH.
64 	 */
65 	for (se = 0; se < gfx_info->max_shader_engines; se++)
66 		for (sh = 0; sh < gfx_info->max_sh_per_se; sh++)
67 			cu_per_sh[se][sh] = hweight32(
68 				cu_info->bitmap[xcc_inst][se][sh]);
69 
70 	/* Symmetrically map cu_mask to all SEs & SHs:
71 	 * For GFX 12.1.0, the following code only looks at a
72 	 * subset of the cu_mask corresponding to the inst parameter.
73 	 * If we have n XCCs under one GPU node
74 	 * cu_mask[0] bit0 -> XCC0 se_mask[0] bit0 (XCC0,SE0,SH0,CU0)
75 	 * cu_mask[0] bit1 -> XCC1 se_mask[0] bit0 (XCC1,SE0,SH0,CU0)
76 	 * ..
77 	 * cu_mask[0] bitn -> XCCn se_mask[0] bit0 (XCCn,SE0,SH0,CU0)
78 	 * cu_mask[0] bit n+1 -> XCC0 se_mask[1] bit0 (XCC0,SE1,SH0,CU0)
79 	 *
80 	 * For example, if there are 6 XCCs under 1 KFD node, this code
81 	 * running for each inst, will look at the bits as:
82 	 * inst, inst + 6, inst + 12...
83 	 *
84 	 * First ensure all CUs are disabled, then enable user specified CUs.
85 	 */
86 	for (i = 0; i < gfx_info->max_shader_engines; i++)
87 		se_mask[i] = 0;
88 
89 	i = inst;
90 	for (cu = 0; cu < 16; cu++) {
91 		for (sh = 0; sh < gfx_info->max_sh_per_se; sh++) {
92 			for (se = 0; se < gfx_info->max_shader_engines; se++) {
93 				if (cu_per_sh[se][sh] > cu) {
94 					if (cu_mask[i / 32] & (1U << (i % 32))) {
95 						if (cu == 8 && sh == 0)
96 							se_mask[se] |= en_mask << 30;
97 						else
98 							se_mask[se] |= en_mask << (cu_inc + sh * 16);
99 					}
100 					i += inc;
101 					if (i >= cu_mask_count)
102 						return;
103 				}
104 			}
105 		}
106 		cu_inc += 2;
107 	}
108 }
109 
110 static void update_cu_mask(struct mqd_manager *mm, void *mqd,
111 			   struct mqd_update_info *minfo, uint32_t inst)
112 {
113 	struct v12_1_compute_mqd *m;
114 	uint32_t se_mask[2] = {0};
115 
116 	if (!minfo || !minfo->cu_mask.ptr)
117 		return;
118 
119 	mqd_symmetrically_map_cu_mask_v12_1(mm,
120 		minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask, inst);
121 
122 	m = get_mqd(mqd);
123 	m->compute_static_thread_mgmt_se0 = se_mask[0];
124 	m->compute_static_thread_mgmt_se1 = se_mask[1];
125 
126 	pr_debug("update cu mask to %#x %#x\n",
127 		m->compute_static_thread_mgmt_se0,
128 		m->compute_static_thread_mgmt_se1);
129 }
130 
131 static void set_priority(struct v12_1_compute_mqd *m, struct queue_properties *q)
132 {
133 	m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
134 	/* m->cp_hqd_queue_priority = q->priority; */
135 }
136 
137 static struct kfd_mem_obj *allocate_mqd(struct mqd_manager *mm,
138 		struct queue_properties *q)
139 {
140 	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
141 	struct kfd_node *node = mm->dev;
142 	struct kfd_mem_obj *mqd_mem_obj;
143 
144 	if (q->type == KFD_QUEUE_TYPE_COMPUTE)
145 		mqd_size *= NUM_XCC(node->xcc_mask);
146 
147 	if (kfd_gtt_sa_allocate(node, mqd_size, &mqd_mem_obj))
148 		return NULL;
149 
150 	return mqd_mem_obj;
151 }
152 
153 static void init_mqd(struct mqd_manager *mm, void **mqd,
154 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
155 			struct queue_properties *q)
156 {
157 	uint64_t addr;
158 	struct v12_1_compute_mqd *m;
159 	u32 mqd_size = AMDGPU_MQD_SIZE_ALIGN(mm->mqd_size);
160 
161 	m = (struct v12_1_compute_mqd *) mqd_mem_obj->cpu_ptr;
162 	addr = mqd_mem_obj->gpu_addr;
163 
164 	memset(m, 0, mqd_size);
165 
166 	m->header = 0xC0310800;
167 	m->compute_pipelinestat_enable = 1;
168 	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
169 	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
170 	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
171 	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
172 	m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF;
173 	m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF;
174 	m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF;
175 	m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF;
176 	m->compute_static_thread_mgmt_se8 = 0xFFFFFFFF;
177 
178 	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
179 			0x63 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
180 
181 	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
182 
183 	m->cp_mqd_base_addr_lo	= lower_32_bits(addr);
184 	m->cp_mqd_base_addr_hi	= upper_32_bits(addr);
185 
186 	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
187 			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
188 			1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
189 
190 	/* Set cp_hqd_hq_status0.c_queue_debug_en to 1 to have the CP set up the
191 	 * DISPATCH_PTR.  This is required for the kfd debugger
192 	 */
193 	m->cp_hqd_hq_status0 = 1 << 14;
194 
195 	if (amdgpu_amdkfd_have_atomics_support(mm->dev->adev))
196 		m->cp_hqd_hq_status0 |= 1 << 29;
197 
198 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
199 		m->cp_hqd_aql_control =
200 			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
201 	}
202 
203 	if (mm->dev->kfd->cwsr_enabled) {
204 		m->cp_hqd_persistent_state |=
205 			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
206 		m->cp_hqd_ctx_save_base_addr_lo =
207 			lower_32_bits(q->ctx_save_restore_area_address);
208 		m->cp_hqd_ctx_save_base_addr_hi =
209 			upper_32_bits(q->ctx_save_restore_area_address);
210 		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
211 		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
212 		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
213 		m->cp_hqd_wg_state_offset = q->ctl_stack_size;
214 	}
215 
216 	*mqd = m;
217 	if (gart_addr)
218 		*gart_addr = addr;
219 	mm->update_mqd(mm, m, q, NULL);
220 }
221 
222 static int load_mqd(struct mqd_manager *mm, void *mqd,
223 			uint32_t pipe_id, uint32_t queue_id,
224 			struct queue_properties *p, struct mm_struct *mms)
225 {
226 	int r = 0;
227 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
228 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
229 
230 	r = mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id,
231 					  (uint32_t __user *)p->write_ptr,
232 					  wptr_shift, 0, mms, 0);
233 	return r;
234 }
235 
236 static void update_mqd(struct mqd_manager *mm, void *mqd,
237 		       struct queue_properties *q,
238 		       struct mqd_update_info *minfo)
239 {
240 	struct v12_1_compute_mqd *m;
241 
242 	m = get_mqd(mqd);
243 
244 	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
245 	m->cp_hqd_pq_control |=
246 			ffs(q->queue_size / sizeof(unsigned int)) - 1 - 1;
247 	m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__UNORD_DISPATCH_MASK;
248 	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
249 
250 	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
251 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
252 
253 	if (q->metadata_queue_size) {
254 		/* On GC 12.1 is 64 DWs which is 4 times size of AQL packet */
255 		if (q->metadata_queue_size == q->queue_size * 4) {
256 			/*
257 			 * User application allocates main queue ring and metadata queue ring
258 			 * with a single allocation. metadata queue ring starts after main
259 			 * queue ring.
260 			 */
261 			m->cp_hqd_kd_base =
262 				lower_32_bits((q->queue_address + q->queue_size) >> 8);
263 			m->cp_hqd_kd_base_hi =
264 				upper_32_bits((q->queue_address + q->queue_size) >> 8);
265 
266 			m->cp_hqd_kd_cntl |= CP_HQD_KD_CNTL__KD_FETCHER_ENABLE_MASK;
267 			/* KD_SIZE = 2 for metadata packet = 64 DWs */
268 			m->cp_hqd_kd_cntl |= 2 << CP_HQD_KD_CNTL__KD_SIZE__SHIFT;
269 		} else {
270 			pr_warn("Invalid metadata ring size, metadata queue will be ignored\n");
271 		}
272 	}
273 
274 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
275 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
276 	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
277 	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
278 
279 	m->cp_hqd_pq_doorbell_control =
280 		q->doorbell_off <<
281 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
282 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
283 			m->cp_hqd_pq_doorbell_control);
284 
285 	m->cp_hqd_ib_control = 1 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT;
286 
287 	/*
288 	 * HW does not clamp this field correctly. Maximum EOP queue size
289 	 * is constrained by per-SE EOP done signal count, which is 8-bit.
290 	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
291 	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
292 	 * is safe, giving a maximum field value of 0xA.
293 	 */
294 	m->cp_hqd_eop_control = min(0xA,
295 		ffs(q->eop_ring_buffer_size / sizeof(unsigned int)) - 1 - 1);
296 	m->cp_hqd_eop_base_addr_lo =
297 			lower_32_bits(q->eop_ring_buffer_address >> 8);
298 	m->cp_hqd_eop_base_addr_hi =
299 			upper_32_bits(q->eop_ring_buffer_address >> 8);
300 
301 	m->cp_hqd_iq_timer = 0;
302 
303 	m->cp_hqd_vmid = q->vmid;
304 
305 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
306 		/* GC 10 removed WPP_CLAMP from PQ Control */
307 		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
308 				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
309 				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT;
310 		m->cp_hqd_pq_doorbell_control |=
311 			1 << CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
312 	}
313 	if (mm->dev->kfd->cwsr_enabled)
314 		m->cp_hqd_ctx_save_control = 0;
315 
316 	set_priority(m, q);
317 
318 	q->is_active = QUEUE_IS_ACTIVE(*q);
319 }
320 
321 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
322 {
323 	return false;
324 }
325 
326 static int get_wave_state(struct mqd_manager *mm, void *mqd,
327 			  struct queue_properties *q,
328 			  void __user *ctl_stack,
329 			  u32 *ctl_stack_used_size,
330 			  u32 *save_area_used_size)
331 {
332 	struct v12_1_compute_mqd *m;
333 	struct mqd_user_context_save_area_header header;
334 
335 	m = get_mqd(mqd);
336 
337 	/* Control stack is written backwards, while workgroup context data
338 	 * is written forwards. Both starts from m->cp_hqd_cntl_stack_size.
339 	 * Current position is at m->cp_hqd_cntl_stack_offset and
340 	 * m->cp_hqd_wg_state_offset, respectively.
341 	 */
342 	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
343 		m->cp_hqd_cntl_stack_offset;
344 	*save_area_used_size = m->cp_hqd_wg_state_offset -
345 		m->cp_hqd_cntl_stack_size;
346 
347 	/* Control stack is not copied to user mode for GFXv12 because
348 	 * it's part of the context save area that is already
349 	 * accessible to user mode
350 	 */
351 	header.control_stack_size = *ctl_stack_used_size;
352 	header.wave_state_size = *save_area_used_size;
353 
354 	header.wave_state_offset = m->cp_hqd_wg_state_offset;
355 	header.control_stack_offset = m->cp_hqd_cntl_stack_offset;
356 
357 	if (copy_to_user(ctl_stack, &header, sizeof(header)))
358 		return -EFAULT;
359 
360 	return 0;
361 }
362 
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 v12_1_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 
377 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
378 		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
379 		struct queue_properties *q)
380 {
381 	struct v12_sdma_mqd *m;
382 
383 	m = (struct v12_sdma_mqd *) mqd_mem_obj->cpu_ptr;
384 
385 	memset(m, 0, PAGE_SIZE);
386 
387 	*mqd = m;
388 	if (gart_addr)
389 		*gart_addr = mqd_mem_obj->gpu_addr;
390 
391 	mm->update_mqd(mm, m, q, NULL);
392 }
393 
394 #define SDMA_RLC_DUMMY_DEFAULT 0xf
395 
396 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
397 		struct queue_properties *q,
398 		struct mqd_update_info *minfo)
399 {
400 	struct v12_sdma_mqd *m;
401 
402 	m = get_sdma_mqd(mqd);
403 	m->sdmax_rlcx_rb_cntl = (ffs(q->queue_size / sizeof(unsigned int)) - 1)
404 		<< SDMA0_SDMA_QUEUE0_RB_CNTL__RB_SIZE__SHIFT |
405 		q->vmid << SDMA0_SDMA_QUEUE0_RB_CNTL__RB_VMID__SHIFT |
406 		1 << SDMA0_SDMA_QUEUE0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
407 		6 << SDMA0_SDMA_QUEUE0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT |
408 		1 << SDMA0_SDMA_QUEUE0_RB_CNTL__MCU_WPTR_POLL_ENABLE__SHIFT;
409 
410 	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
411 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
412 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
413 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
414 	m->sdmax_rlcx_rb_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
415 	m->sdmax_rlcx_rb_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
416 	m->sdmax_rlcx_doorbell_offset =
417 		q->doorbell_off << SDMA0_SDMA_QUEUE0_DOORBELL_OFFSET__OFFSET__SHIFT;
418 
419 	m->sdmax_rlcx_sched_cntl = (amdgpu_sdma_phase_quantum
420 		<< SDMA0_SDMA_QUEUE0_SCHEDULE_CNTL__CONTEXT_QUANTUM__SHIFT)
421 		 & SDMA0_SDMA_QUEUE0_SCHEDULE_CNTL__CONTEXT_QUANTUM_MASK;
422 
423 	m->sdma_engine_id = q->sdma_engine_id;
424 	m->sdma_queue_id = q->sdma_queue_id;
425 
426 	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
427 
428 	/* Allow context switch so we don't cross-process starve with a massive
429 	* command buffer of long-running SDMA commands
430 	* sdmax_rlcx_ib_cntl represent SDMA_QUEUE0_IB_CNTL register
431 	*/
432 	m->sdmax_rlcx_ib_cntl |= SDMA0_SDMA_QUEUE0_IB_CNTL__SWITCH_INSIDE_IB_MASK;
433 
434 	q->is_active = QUEUE_IS_ACTIVE(*q);
435 }
436 
437 static void get_xcc_mqd(struct kfd_mem_obj *mqd_mem_obj,
438 			       struct kfd_mem_obj *xcc_mqd_mem_obj,
439 			       uint64_t offset)
440 {
441 	xcc_mqd_mem_obj->mem = (offset == 0) ?
442 					mqd_mem_obj->mem : NULL;
443 	xcc_mqd_mem_obj->gpu_addr = mqd_mem_obj->gpu_addr + offset;
444 	xcc_mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t)mqd_mem_obj->cpu_ptr
445 						+ offset);
446 }
447 
448 static void init_mqd_v12_1(struct mqd_manager *mm, void **mqd,
449 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
450 			struct queue_properties *q)
451 {
452 	struct v12_1_compute_mqd *m;
453 	int xcc = 0;
454 	struct kfd_mem_obj xcc_mqd_mem_obj;
455 	uint64_t xcc_gart_addr = 0;
456 	uint64_t xcc_ctx_save_restore_area_address;
457 	uint64_t offset = mm->mqd_stride(mm, q);
458 	uint32_t local_xcc_start = mm->dev->dqm->current_logical_xcc_start++;
459 
460 	memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj));
461 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
462 		get_xcc_mqd(mqd_mem_obj, &xcc_mqd_mem_obj, offset*xcc);
463 
464 		init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q);
465 
466 		m->cp_mqd_stride_size = offset;
467 
468 		/*
469 		 * Update the CWSR address for each XCC if CWSR is enabled
470 		 * and CWSR area is allocated in thunk
471 		 */
472 		if (mm->dev->kfd->cwsr_enabled &&
473 		    q->ctx_save_restore_area_address) {
474 			xcc_ctx_save_restore_area_address =
475 				q->ctx_save_restore_area_address +
476 				(xcc * q->ctx_save_restore_area_size);
477 
478 			m->cp_hqd_ctx_save_base_addr_lo =
479 				lower_32_bits(xcc_ctx_save_restore_area_address);
480 			m->cp_hqd_ctx_save_base_addr_hi =
481 				upper_32_bits(xcc_ctx_save_restore_area_address);
482 		}
483 
484 		if (q->format == KFD_QUEUE_FORMAT_AQL) {
485 			m->compute_tg_chunk_size = 1;
486 			m->compute_current_logical_xcc_id =
487 					(local_xcc_start + xcc) %
488 					NUM_XCC(mm->dev->xcc_mask);
489 		} else {
490 			/* PM4 Queue */
491 			m->compute_current_logical_xcc_id = 0;
492 			m->compute_tg_chunk_size = 0;
493 			m->pm4_target_xcc_in_xcp = q->pm4_target_xcc;
494 		}
495 
496 		if (xcc == 0) {
497 			/* Set the MQD pointer and gart address to XCC0 MQD */
498 			*mqd = m;
499 			*gart_addr = xcc_gart_addr;
500 		}
501 	}
502 }
503 
504 static void update_mqd_v12_1(struct mqd_manager *mm, void *mqd,
505 		      struct queue_properties *q, struct mqd_update_info *minfo)
506 {
507 	struct v12_1_compute_mqd *m;
508 	int xcc = 0;
509 	uint64_t size = mm->mqd_stride(mm, q);
510 
511 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
512 		m = get_mqd(mqd + size * xcc);
513 		update_mqd(mm, m, q, minfo);
514 
515 		update_cu_mask(mm, m, minfo, xcc);
516 
517 		if (q->format == KFD_QUEUE_FORMAT_AQL) {
518 			m->compute_tg_chunk_size = 1;
519 		} else {
520 			/* PM4 Queue */
521 			m->compute_current_logical_xcc_id = 0;
522 			m->compute_tg_chunk_size = 0;
523 			m->pm4_target_xcc_in_xcp = q->pm4_target_xcc;
524 		}
525 	}
526 }
527 
528 static int destroy_mqd_v12_1(struct mqd_manager *mm, void *mqd,
529 		   enum kfd_preempt_type type, unsigned int timeout,
530 		   uint32_t pipe_id, uint32_t queue_id)
531 {
532 	uint32_t xcc_mask = mm->dev->xcc_mask;
533 	int xcc_id, err, inst = 0;
534 	void *xcc_mqd;
535 	struct v12_1_compute_mqd *m;
536 	uint64_t mqd_offset;
537 
538 	m = get_mqd(mqd);
539 	mqd_offset = m->cp_mqd_stride_size;
540 
541 	for_each_inst(xcc_id, xcc_mask) {
542 		xcc_mqd = mqd + mqd_offset * inst;
543 		err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd,
544 						    type, timeout, pipe_id,
545 						    queue_id, xcc_id);
546 		if (err) {
547 			pr_debug("Destroy MQD failed for xcc: %d\n", inst);
548 			break;
549 		}
550 		++inst;
551 	}
552 
553 	return err;
554 }
555 
556 static int load_mqd_v12_1(struct mqd_manager *mm, void *mqd,
557 			uint32_t pipe_id, uint32_t queue_id,
558 			struct queue_properties *p, struct mm_struct *mms)
559 {
560 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
561 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
562 	uint32_t xcc_mask = mm->dev->xcc_mask;
563 	int xcc_id, err, inst = 0;
564 	void *xcc_mqd;
565 	uint64_t mqd_stride_size = mm->mqd_stride(mm, p);
566 
567 	for_each_inst(xcc_id, xcc_mask) {
568 		xcc_mqd = mqd + mqd_stride_size * inst;
569 		err = mm->dev->kfd2kgd->hqd_load(
570 			mm->dev->adev, xcc_mqd, pipe_id, queue_id,
571 			(uint32_t __user *)p->write_ptr, wptr_shift, 0, mms,
572 			xcc_id);
573 		if (err) {
574 			pr_debug("Load MQD failed for xcc: %d\n", inst);
575 			break;
576 		}
577 		++inst;
578 	}
579 
580 	return err;
581 }
582 
583 static int get_wave_state_v12_1(struct mqd_manager *mm, void *mqd,
584 				 struct queue_properties *q,
585 				 void __user *ctl_stack,
586 				 u32 *ctl_stack_used_size,
587 				 u32 *save_area_used_size)
588 {
589 	int xcc, err = 0;
590 	void *xcc_mqd;
591 	void __user *xcc_ctl_stack;
592 	uint64_t mqd_stride_size = mm->mqd_stride(mm, q);
593 	u32 tmp_ctl_stack_used_size = 0, tmp_save_area_used_size = 0;
594 
595 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
596 		xcc_mqd = mqd + mqd_stride_size * xcc;
597 		xcc_ctl_stack = (void __user *)((uintptr_t)ctl_stack +
598 					q->ctx_save_restore_area_size * xcc);
599 
600 		err = get_wave_state(mm, xcc_mqd, q, xcc_ctl_stack,
601 				     &tmp_ctl_stack_used_size,
602 				     &tmp_save_area_used_size);
603 		if (err)
604 			break;
605 
606 		/*
607 		 * Set the ctl_stack_used_size and save_area_used_size to
608 		 * ctl_stack_used_size and save_area_used_size of XCC 0 when
609 		 * passing the info to user-space.
610 		 * For multi XCC, user-space would have to look at the header
611 		 * info of each Control stack area to determine the control
612 		 * stack size and save area used.
613 		 */
614 		if (xcc == 0) {
615 			*ctl_stack_used_size = tmp_ctl_stack_used_size;
616 			*save_area_used_size = tmp_save_area_used_size;
617 		}
618 	}
619 
620 	return err;
621 }
622 
623 #if defined(CONFIG_DEBUG_FS)
624 
625 static int debugfs_show_mqd(struct seq_file *m, void *data)
626 {
627 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
628 		     data, sizeof(struct v12_1_compute_mqd), false);
629 	return 0;
630 }
631 
632 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
633 {
634 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
635 		     data, sizeof(struct v12_sdma_mqd), false);
636 	return 0;
637 }
638 
639 #endif
640 
641 struct mqd_manager *mqd_manager_init_v12_1(enum KFD_MQD_TYPE type,
642 		struct kfd_node *dev)
643 {
644 	struct mqd_manager *mqd;
645 
646 	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
647 		return NULL;
648 
649 	mqd = kzalloc_obj(*mqd);
650 	if (!mqd)
651 		return NULL;
652 
653 	mqd->dev = dev;
654 
655 	switch (type) {
656 	case KFD_MQD_TYPE_CP:
657 		pr_debug("%s@%i\n", __func__, __LINE__);
658 		mqd->allocate_mqd = allocate_mqd;
659 		mqd->init_mqd = init_mqd_v12_1;
660 		mqd->free_mqd = kfd_free_mqd_cp;
661 		mqd->load_mqd = load_mqd_v12_1;
662 		mqd->update_mqd = update_mqd_v12_1;
663 		mqd->destroy_mqd = destroy_mqd_v12_1;
664 		mqd->is_occupied = kfd_is_occupied_cp;
665 		mqd->mqd_size = sizeof(struct v12_1_compute_mqd);
666 		mqd->get_wave_state = get_wave_state_v12_1;
667 		mqd->mqd_stride = kfd_mqd_stride;
668 #if defined(CONFIG_DEBUG_FS)
669 		mqd->debugfs_show_mqd = debugfs_show_mqd;
670 #endif
671 		pr_debug("%s@%i\n", __func__, __LINE__);
672 		break;
673 	case KFD_MQD_TYPE_HIQ:
674 		pr_debug("%s@%i\n", __func__, __LINE__);
675 		mqd->allocate_mqd = allocate_hiq_mqd;
676 		mqd->init_mqd = init_mqd_hiq;
677 		mqd->free_mqd = free_mqd_hiq_sdma;
678 		mqd->load_mqd = kfd_hiq_load_mqd_kiq;
679 		mqd->update_mqd = update_mqd;
680 		mqd->destroy_mqd = kfd_destroy_mqd_cp;
681 		mqd->is_occupied = kfd_is_occupied_cp;
682 		mqd->mqd_size = sizeof(struct v12_1_compute_mqd);
683 		mqd->mqd_stride = kfd_mqd_stride;
684 #if defined(CONFIG_DEBUG_FS)
685 		mqd->debugfs_show_mqd = debugfs_show_mqd;
686 #endif
687 		mqd->check_preemption_failed = check_preemption_failed;
688 		pr_debug("%s@%i\n", __func__, __LINE__);
689 		break;
690 	case KFD_MQD_TYPE_DIQ:
691 		mqd->allocate_mqd = allocate_mqd;
692 		mqd->init_mqd = init_mqd_hiq;
693 		mqd->free_mqd = kfd_free_mqd_cp;
694 		mqd->load_mqd = load_mqd;
695 		mqd->update_mqd = update_mqd;
696 		mqd->destroy_mqd = kfd_destroy_mqd_cp;
697 		mqd->is_occupied = kfd_is_occupied_cp;
698 		mqd->mqd_size = sizeof(struct v12_1_compute_mqd);
699 #if defined(CONFIG_DEBUG_FS)
700 		mqd->debugfs_show_mqd = debugfs_show_mqd;
701 #endif
702 		break;
703 	case KFD_MQD_TYPE_SDMA:
704 		pr_debug("%s@%i\n", __func__, __LINE__);
705 		mqd->allocate_mqd = allocate_mqd;
706 		mqd->init_mqd = init_mqd_sdma;
707 		mqd->free_mqd = kfd_free_mqd_cp;
708 		mqd->load_mqd = kfd_load_mqd_sdma;
709 		mqd->update_mqd = update_mqd_sdma;
710 		mqd->destroy_mqd = kfd_destroy_mqd_sdma;
711 		mqd->is_occupied = kfd_is_occupied_sdma;
712 		mqd->mqd_size = sizeof(struct v12_sdma_mqd);
713 		mqd->mqd_stride = kfd_mqd_stride;
714 #if defined(CONFIG_DEBUG_FS)
715 		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
716 #endif
717 		pr_debug("%s@%i\n", __func__, __LINE__);
718 		break;
719 	default:
720 		kfree(mqd);
721 		return NULL;
722 	}
723 
724 	return mqd;
725 }
726