xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c (revision c060f8168bdf22aa986970955af99702d142dfbe)
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 #include "kfd_device_queue_manager.h"
36 
37 static void update_mqd(struct mqd_manager *mm, void *mqd,
38 		       struct queue_properties *q,
39 		       struct mqd_update_info *minfo);
40 
41 static uint64_t mqd_stride_v9(struct mqd_manager *mm,
42 				struct queue_properties *q)
43 {
44 	if (mm->dev->kfd->cwsr_enabled &&
45 	    q->type == KFD_QUEUE_TYPE_COMPUTE)
46 		return ALIGN(q->ctl_stack_size, PAGE_SIZE) +
47 			ALIGN(sizeof(struct v9_mqd), PAGE_SIZE);
48 
49 	return mm->mqd_size;
50 }
51 
52 static inline struct v9_mqd *get_mqd(void *mqd)
53 {
54 	return (struct v9_mqd *)mqd;
55 }
56 
57 static inline struct v9_sdma_mqd *get_sdma_mqd(void *mqd)
58 {
59 	return (struct v9_sdma_mqd *)mqd;
60 }
61 
62 static void update_cu_mask(struct mqd_manager *mm, void *mqd,
63 			struct mqd_update_info *minfo, uint32_t inst)
64 {
65 	struct v9_mqd *m;
66 	uint32_t se_mask[KFD_MAX_NUM_SE] = {0};
67 
68 	if (!minfo || !minfo->cu_mask.ptr)
69 		return;
70 
71 	mqd_symmetrically_map_cu_mask(mm,
72 		minfo->cu_mask.ptr, minfo->cu_mask.count, se_mask, inst);
73 
74 	m = get_mqd(mqd);
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 	if (KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 3) &&
81 	    KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 4)) {
82 		m->compute_static_thread_mgmt_se4 = se_mask[4];
83 		m->compute_static_thread_mgmt_se5 = se_mask[5];
84 		m->compute_static_thread_mgmt_se6 = se_mask[6];
85 		m->compute_static_thread_mgmt_se7 = se_mask[7];
86 
87 		pr_debug("update cu mask to %#x %#x %#x %#x %#x %#x %#x %#x\n",
88 			m->compute_static_thread_mgmt_se0,
89 			m->compute_static_thread_mgmt_se1,
90 			m->compute_static_thread_mgmt_se2,
91 			m->compute_static_thread_mgmt_se3,
92 			m->compute_static_thread_mgmt_se4,
93 			m->compute_static_thread_mgmt_se5,
94 			m->compute_static_thread_mgmt_se6,
95 			m->compute_static_thread_mgmt_se7);
96 	} else {
97 		pr_debug("inst: %u, update cu mask to %#x %#x %#x %#x\n",
98 			inst, m->compute_static_thread_mgmt_se0,
99 			m->compute_static_thread_mgmt_se1,
100 			m->compute_static_thread_mgmt_se2,
101 			m->compute_static_thread_mgmt_se3);
102 	}
103 }
104 
105 static void set_priority(struct v9_mqd *m, struct queue_properties *q)
106 {
107 	m->cp_hqd_pipe_priority = pipe_priority_map[q->priority];
108 	m->cp_hqd_queue_priority = q->priority;
109 }
110 
111 static struct kfd_mem_obj *allocate_mqd(struct kfd_node *node,
112 		struct queue_properties *q)
113 {
114 	int retval;
115 	struct kfd_mem_obj *mqd_mem_obj = NULL;
116 
117 	/* For V9 only, due to a HW bug, the control stack of a user mode
118 	 * compute queue needs to be allocated just behind the page boundary
119 	 * of its regular MQD buffer. So we allocate an enlarged MQD buffer:
120 	 * the first page of the buffer serves as the regular MQD buffer
121 	 * purpose and the remaining is for control stack. Although the two
122 	 * parts are in the same buffer object, they need different memory
123 	 * types: MQD part needs UC (uncached) as usual, while control stack
124 	 * needs NC (non coherent), which is different from the UC type which
125 	 * is used when control stack is allocated in user space.
126 	 *
127 	 * Because of all those, we use the gtt allocation function instead
128 	 * of sub-allocation function for this enlarged MQD buffer. Moreover,
129 	 * in order to achieve two memory types in a single buffer object, we
130 	 * pass a special bo flag AMDGPU_GEM_CREATE_CP_MQD_GFX9 to instruct
131 	 * amdgpu memory functions to do so.
132 	 */
133 	if (node->kfd->cwsr_enabled && (q->type == KFD_QUEUE_TYPE_COMPUTE)) {
134 		mqd_mem_obj = kzalloc(sizeof(struct kfd_mem_obj), GFP_KERNEL);
135 		if (!mqd_mem_obj)
136 			return NULL;
137 		retval = amdgpu_amdkfd_alloc_gtt_mem(node->adev,
138 			(ALIGN(q->ctl_stack_size, PAGE_SIZE) +
139 			ALIGN(sizeof(struct v9_mqd), PAGE_SIZE)) *
140 			NUM_XCC(node->xcc_mask),
141 			&(mqd_mem_obj->gtt_mem),
142 			&(mqd_mem_obj->gpu_addr),
143 			(void *)&(mqd_mem_obj->cpu_ptr), true);
144 
145 		if (retval) {
146 			kfree(mqd_mem_obj);
147 			return NULL;
148 		}
149 	} else {
150 		retval = kfd_gtt_sa_allocate(node, sizeof(struct v9_mqd),
151 				&mqd_mem_obj);
152 		if (retval)
153 			return NULL;
154 	}
155 
156 	return mqd_mem_obj;
157 }
158 
159 static void init_mqd(struct mqd_manager *mm, void **mqd,
160 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
161 			struct queue_properties *q)
162 {
163 	uint64_t addr;
164 	struct v9_mqd *m;
165 
166 	m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr;
167 	addr = mqd_mem_obj->gpu_addr;
168 
169 	memset(m, 0, sizeof(struct v9_mqd));
170 
171 	m->header = 0xC0310800;
172 	m->compute_pipelinestat_enable = 1;
173 	m->compute_static_thread_mgmt_se0 = 0xFFFFFFFF;
174 	m->compute_static_thread_mgmt_se1 = 0xFFFFFFFF;
175 	m->compute_static_thread_mgmt_se2 = 0xFFFFFFFF;
176 	m->compute_static_thread_mgmt_se3 = 0xFFFFFFFF;
177 	m->compute_static_thread_mgmt_se4 = 0xFFFFFFFF;
178 	m->compute_static_thread_mgmt_se5 = 0xFFFFFFFF;
179 	m->compute_static_thread_mgmt_se6 = 0xFFFFFFFF;
180 	m->compute_static_thread_mgmt_se7 = 0xFFFFFFFF;
181 
182 	m->cp_hqd_persistent_state = CP_HQD_PERSISTENT_STATE__PRELOAD_REQ_MASK |
183 			0x53 << CP_HQD_PERSISTENT_STATE__PRELOAD_SIZE__SHIFT;
184 
185 	m->cp_mqd_control = 1 << CP_MQD_CONTROL__PRIV_STATE__SHIFT;
186 
187 	m->cp_mqd_base_addr_lo        = lower_32_bits(addr);
188 	m->cp_mqd_base_addr_hi        = upper_32_bits(addr);
189 
190 	m->cp_hqd_quantum = 1 << CP_HQD_QUANTUM__QUANTUM_EN__SHIFT |
191 			1 << CP_HQD_QUANTUM__QUANTUM_SCALE__SHIFT |
192 			1 << CP_HQD_QUANTUM__QUANTUM_DURATION__SHIFT;
193 
194 	/* Set cp_hqd_hq_scheduler0 bit 14 to 1 to have the CP set up the
195 	 * DISPATCH_PTR.  This is required for the kfd debugger
196 	 */
197 	m->cp_hqd_hq_status0 = 1 << 14;
198 
199 	if (q->format == KFD_QUEUE_FORMAT_AQL)
200 		m->cp_hqd_aql_control =
201 			1 << CP_HQD_AQL_CONTROL__CONTROL0__SHIFT;
202 
203 	if (q->tba_addr) {
204 		m->compute_pgm_rsrc2 |=
205 			(1 << COMPUTE_PGM_RSRC2__TRAP_PRESENT__SHIFT);
206 	}
207 
208 	if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address) {
209 		m->cp_hqd_persistent_state |=
210 			(1 << CP_HQD_PERSISTENT_STATE__QSWITCH_MODE__SHIFT);
211 		m->cp_hqd_ctx_save_base_addr_lo =
212 			lower_32_bits(q->ctx_save_restore_area_address);
213 		m->cp_hqd_ctx_save_base_addr_hi =
214 			upper_32_bits(q->ctx_save_restore_area_address);
215 		m->cp_hqd_ctx_save_size = q->ctx_save_restore_area_size;
216 		m->cp_hqd_cntl_stack_size = q->ctl_stack_size;
217 		m->cp_hqd_cntl_stack_offset = q->ctl_stack_size;
218 		m->cp_hqd_wg_state_offset = q->ctl_stack_size;
219 	}
220 
221 	*mqd = m;
222 	if (gart_addr)
223 		*gart_addr = addr;
224 	update_mqd(mm, m, q, NULL);
225 }
226 
227 static int load_mqd(struct mqd_manager *mm, void *mqd,
228 			uint32_t pipe_id, uint32_t queue_id,
229 			struct queue_properties *p, struct mm_struct *mms)
230 {
231 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
232 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
233 
234 	return mm->dev->kfd2kgd->hqd_load(mm->dev->adev, mqd, pipe_id, queue_id,
235 					  (uint32_t __user *)p->write_ptr,
236 					  wptr_shift, 0, mms, 0);
237 }
238 
239 static void update_mqd(struct mqd_manager *mm, void *mqd,
240 			struct queue_properties *q,
241 			struct mqd_update_info *minfo)
242 {
243 	struct v9_mqd *m;
244 
245 	m = get_mqd(mqd);
246 
247 	m->cp_hqd_pq_control = 5 << CP_HQD_PQ_CONTROL__RPTR_BLOCK_SIZE__SHIFT;
248 	m->cp_hqd_pq_control |= order_base_2(q->queue_size / 4) - 1;
249 	pr_debug("cp_hqd_pq_control 0x%x\n", m->cp_hqd_pq_control);
250 
251 	m->cp_hqd_pq_base_lo = lower_32_bits((uint64_t)q->queue_address >> 8);
252 	m->cp_hqd_pq_base_hi = upper_32_bits((uint64_t)q->queue_address >> 8);
253 
254 	m->cp_hqd_pq_rptr_report_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
255 	m->cp_hqd_pq_rptr_report_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
256 	m->cp_hqd_pq_wptr_poll_addr_lo = lower_32_bits((uint64_t)q->write_ptr);
257 	m->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits((uint64_t)q->write_ptr);
258 
259 	m->cp_hqd_pq_doorbell_control =
260 		q->doorbell_off <<
261 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
262 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
263 			m->cp_hqd_pq_doorbell_control);
264 
265 	m->cp_hqd_ib_control =
266 		3 << CP_HQD_IB_CONTROL__MIN_IB_AVAIL_SIZE__SHIFT |
267 		1 << CP_HQD_IB_CONTROL__IB_EXE_DISABLE__SHIFT;
268 
269 	/*
270 	 * HW does not clamp this field correctly. Maximum EOP queue size
271 	 * is constrained by per-SE EOP done signal count, which is 8-bit.
272 	 * Limit is 0xFF EOP entries (= 0x7F8 dwords). CP will not submit
273 	 * more than (EOP entry count - 1) so a queue size of 0x800 dwords
274 	 * is safe, giving a maximum field value of 0xA.
275 	 *
276 	 * Also, do calculation only if EOP is used (size > 0), otherwise
277 	 * the order_base_2 calculation provides incorrect result.
278 	 *
279 	 */
280 	m->cp_hqd_eop_control = q->eop_ring_buffer_size ?
281 		min(0xA, order_base_2(q->eop_ring_buffer_size / 4) - 1) : 0;
282 
283 	m->cp_hqd_eop_base_addr_lo =
284 			lower_32_bits(q->eop_ring_buffer_address >> 8);
285 	m->cp_hqd_eop_base_addr_hi =
286 			upper_32_bits(q->eop_ring_buffer_address >> 8);
287 
288 	m->cp_hqd_iq_timer = 0;
289 
290 	m->cp_hqd_vmid = q->vmid;
291 
292 	if (q->format == KFD_QUEUE_FORMAT_AQL) {
293 		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
294 				2 << CP_HQD_PQ_CONTROL__SLOT_BASED_WPTR__SHIFT |
295 				1 << CP_HQD_PQ_CONTROL__QUEUE_FULL_EN__SHIFT |
296 				1 << CP_HQD_PQ_CONTROL__WPP_CLAMP_EN__SHIFT;
297 		m->cp_hqd_pq_doorbell_control |= 1 <<
298 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_BIF_DROP__SHIFT;
299 	}
300 	if (mm->dev->kfd->cwsr_enabled && q->ctx_save_restore_area_address)
301 		m->cp_hqd_ctx_save_control = 0;
302 
303 	if (KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 3) &&
304 	    KFD_GC_VERSION(mm->dev) != IP_VERSION(9, 4, 4))
305 		update_cu_mask(mm, mqd, minfo, 0);
306 	set_priority(m, q);
307 
308 	if (minfo && KFD_GC_VERSION(mm->dev) >= IP_VERSION(9, 4, 2)) {
309 		if (minfo->update_flag & UPDATE_FLAG_IS_GWS)
310 			m->compute_resource_limits |=
311 				COMPUTE_RESOURCE_LIMITS__FORCE_SIMD_DIST_MASK;
312 		else
313 			m->compute_resource_limits &=
314 				~COMPUTE_RESOURCE_LIMITS__FORCE_SIMD_DIST_MASK;
315 	}
316 
317 	q->is_active = QUEUE_IS_ACTIVE(*q);
318 }
319 
320 
321 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
322 {
323 	struct v9_mqd *m = (struct v9_mqd *)mqd;
324 	uint32_t doorbell_id = m->queue_doorbell_id0;
325 
326 	m->queue_doorbell_id0 = 0;
327 
328 	return kfd_check_hiq_mqd_doorbell_id(mm->dev, doorbell_id, 0);
329 }
330 
331 static int get_wave_state(struct mqd_manager *mm, void *mqd,
332 			  struct queue_properties *q,
333 			  void __user *ctl_stack,
334 			  u32 *ctl_stack_used_size,
335 			  u32 *save_area_used_size)
336 {
337 	struct v9_mqd *m;
338 	struct kfd_context_save_area_header header;
339 
340 	/* Control stack is located one page after MQD. */
341 	void *mqd_ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE);
342 
343 	m = get_mqd(mqd);
344 
345 	*ctl_stack_used_size = m->cp_hqd_cntl_stack_size -
346 		m->cp_hqd_cntl_stack_offset;
347 	*save_area_used_size = m->cp_hqd_wg_state_offset -
348 		m->cp_hqd_cntl_stack_size;
349 
350 	header.wave_state.control_stack_size = *ctl_stack_used_size;
351 	header.wave_state.wave_state_size = *save_area_used_size;
352 
353 	header.wave_state.wave_state_offset = m->cp_hqd_wg_state_offset;
354 	header.wave_state.control_stack_offset = m->cp_hqd_cntl_stack_offset;
355 
356 	if (copy_to_user(ctl_stack, &header, sizeof(header.wave_state)))
357 		return -EFAULT;
358 
359 	if (copy_to_user(ctl_stack + m->cp_hqd_cntl_stack_offset,
360 				mqd_ctl_stack + m->cp_hqd_cntl_stack_offset,
361 				*ctl_stack_used_size))
362 		return -EFAULT;
363 
364 	return 0;
365 }
366 
367 static void get_checkpoint_info(struct mqd_manager *mm, void *mqd, u32 *ctl_stack_size)
368 {
369 	struct v9_mqd *m = get_mqd(mqd);
370 
371 	*ctl_stack_size = m->cp_hqd_cntl_stack_size;
372 }
373 
374 static void checkpoint_mqd(struct mqd_manager *mm, void *mqd, void *mqd_dst, void *ctl_stack_dst)
375 {
376 	struct v9_mqd *m;
377 	/* Control stack is located one page after MQD. */
378 	void *ctl_stack = (void *)((uintptr_t)mqd + PAGE_SIZE);
379 
380 	m = get_mqd(mqd);
381 
382 	memcpy(mqd_dst, m, sizeof(struct v9_mqd));
383 	memcpy(ctl_stack_dst, ctl_stack, m->cp_hqd_cntl_stack_size);
384 }
385 
386 static void restore_mqd(struct mqd_manager *mm, void **mqd,
387 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
388 			struct queue_properties *qp,
389 			const void *mqd_src,
390 			const void *ctl_stack_src, u32 ctl_stack_size)
391 {
392 	uint64_t addr;
393 	struct v9_mqd *m;
394 	void *ctl_stack;
395 
396 	m = (struct v9_mqd *) mqd_mem_obj->cpu_ptr;
397 	addr = mqd_mem_obj->gpu_addr;
398 
399 	memcpy(m, mqd_src, sizeof(*m));
400 
401 	*mqd = m;
402 	if (gart_addr)
403 		*gart_addr = addr;
404 
405 	/* Control stack is located one page after MQD. */
406 	ctl_stack = (void *)((uintptr_t)*mqd + PAGE_SIZE);
407 	memcpy(ctl_stack, ctl_stack_src, ctl_stack_size);
408 
409 	m->cp_hqd_pq_doorbell_control =
410 		qp->doorbell_off <<
411 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
412 	pr_debug("cp_hqd_pq_doorbell_control 0x%x\n",
413 				m->cp_hqd_pq_doorbell_control);
414 
415 	qp->is_active = 0;
416 }
417 
418 static void init_mqd_hiq(struct mqd_manager *mm, void **mqd,
419 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
420 			struct queue_properties *q)
421 {
422 	struct v9_mqd *m;
423 
424 	init_mqd(mm, mqd, mqd_mem_obj, gart_addr, q);
425 
426 	m = get_mqd(*mqd);
427 
428 	m->cp_hqd_pq_control |= 1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
429 			1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
430 }
431 
432 static int destroy_hiq_mqd(struct mqd_manager *mm, void *mqd,
433 			enum kfd_preempt_type type, unsigned int timeout,
434 			uint32_t pipe_id, uint32_t queue_id)
435 {
436 	int err;
437 	struct v9_mqd *m;
438 	u32 doorbell_off;
439 
440 	m = get_mqd(mqd);
441 
442 	doorbell_off = m->cp_hqd_pq_doorbell_control >>
443 			CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
444 	err = amdgpu_amdkfd_unmap_hiq(mm->dev->adev, doorbell_off, 0);
445 	if (err)
446 		pr_debug("Destroy HIQ MQD failed: %d\n", err);
447 
448 	return err;
449 }
450 
451 static void init_mqd_sdma(struct mqd_manager *mm, void **mqd,
452 		struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
453 		struct queue_properties *q)
454 {
455 	struct v9_sdma_mqd *m;
456 
457 	m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr;
458 
459 	memset(m, 0, sizeof(struct v9_sdma_mqd));
460 
461 	*mqd = m;
462 	if (gart_addr)
463 		*gart_addr = mqd_mem_obj->gpu_addr;
464 
465 	mm->update_mqd(mm, m, q, NULL);
466 }
467 
468 #define SDMA_RLC_DUMMY_DEFAULT 0xf
469 
470 static void update_mqd_sdma(struct mqd_manager *mm, void *mqd,
471 			struct queue_properties *q,
472 			struct mqd_update_info *minfo)
473 {
474 	struct v9_sdma_mqd *m;
475 
476 	m = get_sdma_mqd(mqd);
477 	m->sdmax_rlcx_rb_cntl = order_base_2(q->queue_size / 4)
478 		<< SDMA0_RLC0_RB_CNTL__RB_SIZE__SHIFT |
479 		q->vmid << SDMA0_RLC0_RB_CNTL__RB_VMID__SHIFT |
480 		1 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_ENABLE__SHIFT |
481 		6 << SDMA0_RLC0_RB_CNTL__RPTR_WRITEBACK_TIMER__SHIFT;
482 
483 	m->sdmax_rlcx_rb_base = lower_32_bits(q->queue_address >> 8);
484 	m->sdmax_rlcx_rb_base_hi = upper_32_bits(q->queue_address >> 8);
485 	m->sdmax_rlcx_rb_rptr_addr_lo = lower_32_bits((uint64_t)q->read_ptr);
486 	m->sdmax_rlcx_rb_rptr_addr_hi = upper_32_bits((uint64_t)q->read_ptr);
487 	m->sdmax_rlcx_doorbell_offset =
488 		q->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
489 
490 	m->sdma_engine_id = q->sdma_engine_id;
491 	m->sdma_queue_id = q->sdma_queue_id;
492 	m->sdmax_rlcx_dummy_reg = SDMA_RLC_DUMMY_DEFAULT;
493 
494 	q->is_active = QUEUE_IS_ACTIVE(*q);
495 }
496 
497 static void checkpoint_mqd_sdma(struct mqd_manager *mm,
498 				void *mqd,
499 				void *mqd_dst,
500 				void *ctl_stack_dst)
501 {
502 	struct v9_sdma_mqd *m;
503 
504 	m = get_sdma_mqd(mqd);
505 
506 	memcpy(mqd_dst, m, sizeof(struct v9_sdma_mqd));
507 }
508 
509 static void restore_mqd_sdma(struct mqd_manager *mm, void **mqd,
510 			     struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
511 			     struct queue_properties *qp,
512 			     const void *mqd_src,
513 			     const void *ctl_stack_src, const u32 ctl_stack_size)
514 {
515 	uint64_t addr;
516 	struct v9_sdma_mqd *m;
517 
518 	m = (struct v9_sdma_mqd *) mqd_mem_obj->cpu_ptr;
519 	addr = mqd_mem_obj->gpu_addr;
520 
521 	memcpy(m, mqd_src, sizeof(*m));
522 
523 	m->sdmax_rlcx_doorbell_offset =
524 		qp->doorbell_off << SDMA0_RLC0_DOORBELL_OFFSET__OFFSET__SHIFT;
525 
526 	*mqd = m;
527 	if (gart_addr)
528 		*gart_addr = addr;
529 
530 	qp->is_active = 0;
531 }
532 
533 static void init_mqd_hiq_v9_4_3(struct mqd_manager *mm, void **mqd,
534 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
535 			struct queue_properties *q)
536 {
537 	struct v9_mqd *m;
538 	int xcc = 0;
539 	struct kfd_mem_obj xcc_mqd_mem_obj;
540 	uint64_t xcc_gart_addr = 0;
541 
542 	memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj));
543 
544 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
545 		kfd_get_hiq_xcc_mqd(mm->dev, &xcc_mqd_mem_obj, xcc);
546 
547 		init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q);
548 
549 		m->cp_hqd_pq_control |= CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK |
550 					1 << CP_HQD_PQ_CONTROL__PRIV_STATE__SHIFT |
551 					1 << CP_HQD_PQ_CONTROL__KMD_QUEUE__SHIFT;
552 		if (amdgpu_sriov_vf(mm->dev->adev))
553 			m->cp_hqd_pq_doorbell_control |= 1 <<
554 				CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_MODE__SHIFT;
555 		m->cp_mqd_stride_size = kfd_hiq_mqd_stride(mm->dev);
556 		if (xcc == 0) {
557 			/* Set no_update_rptr = 0 in Master XCC */
558 			m->cp_hqd_pq_control &= ~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
559 
560 			/* Set the MQD pointer and gart address to XCC0 MQD */
561 			*mqd = m;
562 			*gart_addr = xcc_gart_addr;
563 		}
564 	}
565 }
566 
567 static int hiq_load_mqd_kiq_v9_4_3(struct mqd_manager *mm, void *mqd,
568 			uint32_t pipe_id, uint32_t queue_id,
569 			struct queue_properties *p, struct mm_struct *mms)
570 {
571 	uint32_t xcc_mask = mm->dev->xcc_mask;
572 	int xcc_id, err, inst = 0;
573 	void *xcc_mqd;
574 	uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev);
575 
576 	for_each_inst(xcc_id, xcc_mask) {
577 		xcc_mqd = mqd + hiq_mqd_size * inst;
578 		err = mm->dev->kfd2kgd->hiq_mqd_load(mm->dev->adev, xcc_mqd,
579 						     pipe_id, queue_id,
580 						     p->doorbell_off, xcc_id);
581 		if (err) {
582 			pr_debug("Failed to load HIQ MQD for XCC: %d\n", inst);
583 			break;
584 		}
585 		++inst;
586 	}
587 
588 	return err;
589 }
590 
591 static int destroy_hiq_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
592 			enum kfd_preempt_type type, unsigned int timeout,
593 			uint32_t pipe_id, uint32_t queue_id)
594 {
595 	uint32_t xcc_mask = mm->dev->xcc_mask;
596 	int xcc_id, err, inst = 0;
597 	uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev);
598 	struct v9_mqd *m;
599 	u32 doorbell_off;
600 
601 	for_each_inst(xcc_id, xcc_mask) {
602 		m = get_mqd(mqd + hiq_mqd_size * inst);
603 
604 		doorbell_off = m->cp_hqd_pq_doorbell_control >>
605 				CP_HQD_PQ_DOORBELL_CONTROL__DOORBELL_OFFSET__SHIFT;
606 
607 		err = amdgpu_amdkfd_unmap_hiq(mm->dev->adev, doorbell_off, xcc_id);
608 		if (err) {
609 			pr_debug("Destroy HIQ MQD failed for xcc: %d\n", inst);
610 			break;
611 		}
612 		++inst;
613 	}
614 
615 	return err;
616 }
617 
618 static bool check_preemption_failed_v9_4_3(struct mqd_manager *mm, void *mqd)
619 {
620 	uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev);
621 	uint32_t xcc_mask = mm->dev->xcc_mask;
622 	int inst = 0, xcc_id;
623 	struct v9_mqd *m;
624 	bool ret = false;
625 
626 	for_each_inst(xcc_id, xcc_mask) {
627 		m = get_mqd(mqd + hiq_mqd_size * inst);
628 		ret |= kfd_check_hiq_mqd_doorbell_id(mm->dev,
629 					m->queue_doorbell_id0, inst);
630 		m->queue_doorbell_id0 = 0;
631 		++inst;
632 	}
633 
634 	return ret;
635 }
636 
637 static void get_xcc_mqd(struct kfd_mem_obj *mqd_mem_obj,
638 			       struct kfd_mem_obj *xcc_mqd_mem_obj,
639 			       uint64_t offset)
640 {
641 	xcc_mqd_mem_obj->gtt_mem = (offset == 0) ?
642 					mqd_mem_obj->gtt_mem : NULL;
643 	xcc_mqd_mem_obj->gpu_addr = mqd_mem_obj->gpu_addr + offset;
644 	xcc_mqd_mem_obj->cpu_ptr = (uint32_t *)((uintptr_t)mqd_mem_obj->cpu_ptr
645 						+ offset);
646 }
647 
648 static void init_mqd_v9_4_3(struct mqd_manager *mm, void **mqd,
649 			struct kfd_mem_obj *mqd_mem_obj, uint64_t *gart_addr,
650 			struct queue_properties *q)
651 {
652 	struct v9_mqd *m;
653 	int xcc = 0;
654 	struct kfd_mem_obj xcc_mqd_mem_obj;
655 	uint64_t xcc_gart_addr = 0;
656 	uint64_t xcc_ctx_save_restore_area_address;
657 	uint64_t offset = mm->mqd_stride(mm, q);
658 	uint32_t local_xcc_start = mm->dev->dqm->current_logical_xcc_start++;
659 
660 	memset(&xcc_mqd_mem_obj, 0x0, sizeof(struct kfd_mem_obj));
661 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
662 		get_xcc_mqd(mqd_mem_obj, &xcc_mqd_mem_obj, offset*xcc);
663 
664 		init_mqd(mm, (void **)&m, &xcc_mqd_mem_obj, &xcc_gart_addr, q);
665 
666 		m->cp_mqd_stride_size = offset;
667 
668 		/*
669 		 * Update the CWSR address for each XCC if CWSR is enabled
670 		 * and CWSR area is allocated in thunk
671 		 */
672 		if (mm->dev->kfd->cwsr_enabled &&
673 		    q->ctx_save_restore_area_address) {
674 			xcc_ctx_save_restore_area_address =
675 				q->ctx_save_restore_area_address +
676 				(xcc * q->ctx_save_restore_area_size);
677 
678 			m->cp_hqd_ctx_save_base_addr_lo =
679 				lower_32_bits(xcc_ctx_save_restore_area_address);
680 			m->cp_hqd_ctx_save_base_addr_hi =
681 				upper_32_bits(xcc_ctx_save_restore_area_address);
682 		}
683 
684 		if (q->format == KFD_QUEUE_FORMAT_AQL) {
685 			m->compute_tg_chunk_size = 1;
686 			m->compute_current_logic_xcc_id =
687 					(local_xcc_start + xcc) %
688 					NUM_XCC(mm->dev->xcc_mask);
689 
690 			switch (xcc) {
691 			case 0:
692 				/* Master XCC */
693 				m->cp_hqd_pq_control &=
694 					~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
695 				break;
696 			default:
697 				break;
698 			}
699 		} else {
700 			/* PM4 Queue */
701 			m->compute_current_logic_xcc_id = 0;
702 			m->compute_tg_chunk_size = 0;
703 			m->pm4_target_xcc_in_xcp = q->pm4_target_xcc;
704 		}
705 
706 		if (xcc == 0) {
707 			/* Set the MQD pointer and gart address to XCC0 MQD */
708 			*mqd = m;
709 			*gart_addr = xcc_gart_addr;
710 		}
711 	}
712 }
713 
714 static void update_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
715 		      struct queue_properties *q, struct mqd_update_info *minfo)
716 {
717 	struct v9_mqd *m;
718 	int xcc = 0;
719 	uint64_t size = mm->mqd_stride(mm, q);
720 
721 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
722 		m = get_mqd(mqd + size * xcc);
723 		update_mqd(mm, m, q, minfo);
724 
725 		update_cu_mask(mm, m, minfo, xcc);
726 
727 		if (q->format == KFD_QUEUE_FORMAT_AQL) {
728 			switch (xcc) {
729 			case 0:
730 				/* Master XCC */
731 				m->cp_hqd_pq_control &=
732 					~CP_HQD_PQ_CONTROL__NO_UPDATE_RPTR_MASK;
733 				break;
734 			default:
735 				break;
736 			}
737 			m->compute_tg_chunk_size = 1;
738 		} else {
739 			/* PM4 Queue */
740 			m->compute_current_logic_xcc_id = 0;
741 			m->compute_tg_chunk_size = 0;
742 			m->pm4_target_xcc_in_xcp = q->pm4_target_xcc;
743 		}
744 	}
745 }
746 
747 static int destroy_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
748 		   enum kfd_preempt_type type, unsigned int timeout,
749 		   uint32_t pipe_id, uint32_t queue_id)
750 {
751 	uint32_t xcc_mask = mm->dev->xcc_mask;
752 	int xcc_id, err, inst = 0;
753 	void *xcc_mqd;
754 	struct v9_mqd *m;
755 	uint64_t mqd_offset;
756 
757 	m = get_mqd(mqd);
758 	mqd_offset = m->cp_mqd_stride_size;
759 
760 	for_each_inst(xcc_id, xcc_mask) {
761 		xcc_mqd = mqd + mqd_offset * inst;
762 		err = mm->dev->kfd2kgd->hqd_destroy(mm->dev->adev, xcc_mqd,
763 						    type, timeout, pipe_id,
764 						    queue_id, xcc_id);
765 		if (err) {
766 			pr_debug("Destroy MQD failed for xcc: %d\n", inst);
767 			break;
768 		}
769 		++inst;
770 	}
771 
772 	return err;
773 }
774 
775 static int load_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
776 			uint32_t pipe_id, uint32_t queue_id,
777 			struct queue_properties *p, struct mm_struct *mms)
778 {
779 	/* AQL write pointer counts in 64B packets, PM4/CP counts in dwords. */
780 	uint32_t wptr_shift = (p->format == KFD_QUEUE_FORMAT_AQL ? 4 : 0);
781 	uint32_t xcc_mask = mm->dev->xcc_mask;
782 	int xcc_id, err, inst = 0;
783 	void *xcc_mqd;
784 	uint64_t mqd_stride_size = mm->mqd_stride(mm, p);
785 
786 	for_each_inst(xcc_id, xcc_mask) {
787 		xcc_mqd = mqd + mqd_stride_size * inst;
788 		err = mm->dev->kfd2kgd->hqd_load(
789 			mm->dev->adev, xcc_mqd, pipe_id, queue_id,
790 			(uint32_t __user *)p->write_ptr, wptr_shift, 0, mms,
791 			xcc_id);
792 		if (err) {
793 			pr_debug("Load MQD failed for xcc: %d\n", inst);
794 			break;
795 		}
796 		++inst;
797 	}
798 
799 	return err;
800 }
801 
802 static int get_wave_state_v9_4_3(struct mqd_manager *mm, void *mqd,
803 				 struct queue_properties *q,
804 				 void __user *ctl_stack,
805 				 u32 *ctl_stack_used_size,
806 				 u32 *save_area_used_size)
807 {
808 	int xcc, err = 0;
809 	void *xcc_mqd;
810 	void __user *xcc_ctl_stack;
811 	uint64_t mqd_stride_size = mm->mqd_stride(mm, q);
812 	u32 tmp_ctl_stack_used_size = 0, tmp_save_area_used_size = 0;
813 
814 	for (xcc = 0; xcc < NUM_XCC(mm->dev->xcc_mask); xcc++) {
815 		xcc_mqd = mqd + mqd_stride_size * xcc;
816 		xcc_ctl_stack = (void __user *)((uintptr_t)ctl_stack +
817 					q->ctx_save_restore_area_size * xcc);
818 
819 		err = get_wave_state(mm, xcc_mqd, q, xcc_ctl_stack,
820 				     &tmp_ctl_stack_used_size,
821 				     &tmp_save_area_used_size);
822 		if (err)
823 			break;
824 
825 		/*
826 		 * Set the ctl_stack_used_size and save_area_used_size to
827 		 * ctl_stack_used_size and save_area_used_size of XCC 0 when
828 		 * passing the info the user-space.
829 		 * For multi XCC, user-space would have to look at the header
830 		 * info of each Control stack area to determine the control
831 		 * stack size and save area used.
832 		 */
833 		if (xcc == 0) {
834 			*ctl_stack_used_size = tmp_ctl_stack_used_size;
835 			*save_area_used_size = tmp_save_area_used_size;
836 		}
837 	}
838 
839 	return err;
840 }
841 
842 #if defined(CONFIG_DEBUG_FS)
843 
844 static int debugfs_show_mqd(struct seq_file *m, void *data)
845 {
846 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
847 		     data, sizeof(struct v9_mqd), false);
848 	return 0;
849 }
850 
851 static int debugfs_show_mqd_sdma(struct seq_file *m, void *data)
852 {
853 	seq_hex_dump(m, "    ", DUMP_PREFIX_OFFSET, 32, 4,
854 		     data, sizeof(struct v9_sdma_mqd), false);
855 	return 0;
856 }
857 
858 #endif
859 
860 struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
861 		struct kfd_node *dev)
862 {
863 	struct mqd_manager *mqd;
864 
865 	if (WARN_ON(type >= KFD_MQD_TYPE_MAX))
866 		return NULL;
867 
868 	mqd = kzalloc(sizeof(*mqd), GFP_KERNEL);
869 	if (!mqd)
870 		return NULL;
871 
872 	mqd->dev = dev;
873 
874 	switch (type) {
875 	case KFD_MQD_TYPE_CP:
876 		mqd->allocate_mqd = allocate_mqd;
877 		mqd->free_mqd = kfd_free_mqd_cp;
878 		mqd->is_occupied = kfd_is_occupied_cp;
879 		mqd->get_checkpoint_info = get_checkpoint_info;
880 		mqd->checkpoint_mqd = checkpoint_mqd;
881 		mqd->restore_mqd = restore_mqd;
882 		mqd->mqd_size = sizeof(struct v9_mqd);
883 		mqd->mqd_stride = mqd_stride_v9;
884 #if defined(CONFIG_DEBUG_FS)
885 		mqd->debugfs_show_mqd = debugfs_show_mqd;
886 #endif
887 		if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3) ||
888 		    KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)) {
889 			mqd->init_mqd = init_mqd_v9_4_3;
890 			mqd->load_mqd = load_mqd_v9_4_3;
891 			mqd->update_mqd = update_mqd_v9_4_3;
892 			mqd->destroy_mqd = destroy_mqd_v9_4_3;
893 			mqd->get_wave_state = get_wave_state_v9_4_3;
894 		} else {
895 			mqd->init_mqd = init_mqd;
896 			mqd->load_mqd = load_mqd;
897 			mqd->update_mqd = update_mqd;
898 			mqd->destroy_mqd = kfd_destroy_mqd_cp;
899 			mqd->get_wave_state = get_wave_state;
900 		}
901 		break;
902 	case KFD_MQD_TYPE_HIQ:
903 		mqd->allocate_mqd = allocate_hiq_mqd;
904 		mqd->free_mqd = free_mqd_hiq_sdma;
905 		mqd->update_mqd = update_mqd;
906 		mqd->is_occupied = kfd_is_occupied_cp;
907 		mqd->mqd_size = sizeof(struct v9_mqd);
908 		mqd->mqd_stride = kfd_mqd_stride;
909 #if defined(CONFIG_DEBUG_FS)
910 		mqd->debugfs_show_mqd = debugfs_show_mqd;
911 #endif
912 		if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3) ||
913 		    KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 4)) {
914 			mqd->init_mqd = init_mqd_hiq_v9_4_3;
915 			mqd->load_mqd = hiq_load_mqd_kiq_v9_4_3;
916 			mqd->destroy_mqd = destroy_hiq_mqd_v9_4_3;
917 			mqd->check_preemption_failed = check_preemption_failed_v9_4_3;
918 		} else {
919 			mqd->init_mqd = init_mqd_hiq;
920 			mqd->load_mqd = kfd_hiq_load_mqd_kiq;
921 			mqd->destroy_mqd = destroy_hiq_mqd;
922 			mqd->check_preemption_failed = check_preemption_failed;
923 		}
924 		break;
925 	case KFD_MQD_TYPE_DIQ:
926 		mqd->allocate_mqd = allocate_mqd;
927 		mqd->init_mqd = init_mqd_hiq;
928 		mqd->free_mqd = kfd_free_mqd_cp;
929 		mqd->load_mqd = load_mqd;
930 		mqd->update_mqd = update_mqd;
931 		mqd->destroy_mqd = kfd_destroy_mqd_cp;
932 		mqd->is_occupied = kfd_is_occupied_cp;
933 		mqd->mqd_size = sizeof(struct v9_mqd);
934 #if defined(CONFIG_DEBUG_FS)
935 		mqd->debugfs_show_mqd = debugfs_show_mqd;
936 #endif
937 		break;
938 	case KFD_MQD_TYPE_SDMA:
939 		mqd->allocate_mqd = allocate_sdma_mqd;
940 		mqd->init_mqd = init_mqd_sdma;
941 		mqd->free_mqd = free_mqd_hiq_sdma;
942 		mqd->load_mqd = kfd_load_mqd_sdma;
943 		mqd->update_mqd = update_mqd_sdma;
944 		mqd->destroy_mqd = kfd_destroy_mqd_sdma;
945 		mqd->is_occupied = kfd_is_occupied_sdma;
946 		mqd->checkpoint_mqd = checkpoint_mqd_sdma;
947 		mqd->restore_mqd = restore_mqd_sdma;
948 		mqd->mqd_size = sizeof(struct v9_sdma_mqd);
949 		mqd->mqd_stride = kfd_mqd_stride;
950 #if defined(CONFIG_DEBUG_FS)
951 		mqd->debugfs_show_mqd = debugfs_show_mqd_sdma;
952 #endif
953 		break;
954 	default:
955 		kfree(mqd);
956 		return NULL;
957 	}
958 
959 	return mqd;
960 }
961