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